Download Real business cycles: A Reader

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Predictive analytics wikipedia , lookup

Sensitivity analysis wikipedia , lookup

Transcript
Real business cycles
The Real Business Cycle model has become the dominant mode of business
analysis within the new classical school of macroeconomic thought. It
has been the focus of a great deal of debate and controversy, and yet, to
date, there has been no single source for material on real business cycles,
much less one which provides a balanced account of both sides of the
debate.
This volume fills the gap by presenting a collection of the essential
articles that define and develop the real business cycle school, as well as
those that criticize it. Key areas covered include:
•
•
•
•
•
the establishment of the real business cycle program
the aims and methods of the real business cycle school
analysis of the statistics and econometrics of the calibration techniques
advocated by real businesss cycle modelers
assessment of the empirical success of the real business cycle model
from a variety of methods and perspectives
the measurement of technology shocks in business cycle models (the
Solow residual).
A detailed Introduction assesses the strengths and weaknesses of real
business cycle theory from a critical perspective and includes a nontechnical User’s Guide to the formulation and solution of models which
will aid understanding of the articles and issues discussed.
Offering a thorough assessment of the real business cycle program, this
volume will be a vital resource for students and professionals in
macroeconomics.
James E.Hartley is Assistant Professor of Economics at Mount Holyoke
College, Massachusetts. Kevin D.Hoover is Professor of Economics at the
University of California, Davis. Kevin D.Salyer is Associate Professor of
Economics, also at the University of California, Davis.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
Real business cycles
A Reader
Edited by
James E.Hartley,
Kevin D.Hoover,
and
Kevin D.Salyer
London and New York
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
First published 1998 by Routledge
11 New Fetter Lane, London EC4P 4EE
This edition published in the Taylor & Francis e-Library, 2006.
“To purchase your own copy of this or any of Taylor & Francis or Routledge’s
collection of thousands of eBooks please go to www.eBookstore.tandf.co.uk.”
Simultaneously published in the USA and Canada
by Routledge
29 West 35th Street, New York, NY 10001
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
All rights reserved. No part of this book may be reprinted or
reproduced or utilized in any form or by any electronic,
mechanical, or other means, now known or hereafter
invented, including photocopying and recording, or in any
information storage or retrieval system, without permission
in writing from the publishers.
British Library Cataloguing in Publication Data
A catalogue record for this book is available from the British Library
Library of Congress Cataloging in Publication Data
Real business cycles/[edited by] James E.Hartley, Kevin D.Hoover,
and Kevin D.Salyer.
p.
cm.
Includes bibliographical references and index.
1. Business cycles. I. Hartley, James E., 1966– . I I. Hoover,
Kevin D., 1955– . III. Salyer, Kevin D., 1954– .
H B3711.R35 1998
97–40397
CI P
I S BN 0-203-07071-2 Master e-book I S B N
ISBN 0-203-22306-3 (Adobe eReader Format)
ISBN 0-415-16568-7 (hbk)
IS BN 0-415-17154-7 (pbk)
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
“It is thus that the existence of a common standard of judgment
leads physicists, who are no more saintly than economists, to question
their own best work.”
Steven Weinberg
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
Contents
Acknowledgements
xi
Part I Introduction
1
The Limits of Business Cycle Research
2
A User’s Guide to Solving Real Business Cycle Models
3
43
Part II The foundations of real business cycle modeling
3
4
5
Finn E.Kydland and Edward C.Prescott, “Time to build
and aggregate fluctuations,” Econometrica 50(6), November
1982, pp. 1345–1369.
57
Edward C.Prescott, “Theory ahead of business cycle
measurement,” Federal Reserve Bank of Minneapolis Quarterly
Review 10(4), Fall 1986, pp. 9–22.
83
Lawrence H.Summers, “Some skeptical observations on
real business cycle theory,” Federal Reserve Bank of Minneapolis
Quarterly Review 10(4), Fall 1986, pp. 23–27.
97
6
Edward C.Prescott, “Response to a skeptic,” Federal Reserve
Bank of Minneapolis Quarterly Review 10(4), Fall 1986, pp. 28–33. 102
7
Robert G.King, Charles I.Plosser and Sergio T.Rebelo,
“Production, growth, and business cycles I: The basic
neoclassical model,” Journal of Monetary Economics 21(2), March
1988, pp. 195–232.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
108
viii
CONTENTS
Part III Some extensions
8 Gary D.Hansen, “Indivisible labor and the business cycle,”
Journal of Monetary Economics 16(3), November 1985, pp.
309–328.
149
9 Gary D.Hansen and Randall Wright, “The labor market in
real business cycle theory,” Federal Reserve Bank of Minneapolis
Quarterly Review, Spring 1992, pp. 2–12.
168
10 Lawrence J.Christiano, and Martin Eichenbaum, “Current
real business cycle theories and aggregate labor market
fluctuations,” American Economic Review 82(3), June 1992,
430–450.
179
11 Thomas F.Cooley and Gary D.Hansen, “The inflation tax
in a real business cycle model,” American Economic Review 79(4),
September 1989, pp. 733–748.
200
Part IV The methodology of equilibrium business cycle models
12 Finn E.Kydland and Edward C.Prescott, “The econometrics
of the general equilibrium approach to business cycles,”
Scandinavian Journal of Economics 93(2), 1991, pp. 161–178.
219
13 Finn E.Kydland and Edward C.Prescott, “The computational
experiment: An econometric tool,” Journal of Economic
Perspectives 10(1), Winter 1996, pp. 69–86.
237
14 Lars Peter Hansen and James J.Heckman, “The empirical
foundations of calibration,” Journal of Economic Perspectives
10(1), Winter 1996, pp. 87–104.
254
15 Kevin D.Hoover, “Facts and artifacts: Calibration and the
empirical assessment of real-business-cycle models,” Oxford
Economic Papers 47(1), March 1995, pp. 24–44.
272
Part V The critique of calibration methods
16 Allan W.Gregory and Gregor W.Smith, “Calibration as testing:
Inference in simulated macroeconomic models,” Journal of
Business and Economic Statistics, 9(3), July 1991, pp. 297–303.
295
17 Mark W.Watson, “Measures of fit for calibrated models,”
Journal of Political Economy 101(6), December 1993, pp.
1011–1041.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
302
CONTENTS
ix
18 Fabio Canova, “Statistical inference in calibrated models,”
Journal of Applied Econometrics 9, 1994, pp. 123–144.
333
19 Fabio Canova, “Sensitivity analysis and model evaluation in
simulated dynamic general equilibrium economies,”
International Economic Review 36(2), May 1995, pp. 477–501.
355
Part VI Testing the real business cycle model
20 Finn E.Kydland and Edward C.Prescott. “Business cycles:
Real facts and a monetary myth,” Federal Reserve Bank of
Minneapolis Quarterly Review 14(2), Spring 1990, pp. 3–18.
383
21 Sumru Altuǧ, “Time-to-build and aggregate fluctuations:
Some new evidence,” International Economic Review 30(4),
November 1989, pp. 889–920.
399
22 Fabio Canova, M.Finn, and A.R.Pagan, “Evaluating a real
business cycle model,” in C.Hargreaves (ed.) Nonstationary
Time Series Analysis and Cointegration. Oxford: Oxford University
Press, 1994, pp. 225–255.
431
23 Robert G.King and Charles I.Plosser. “Real business cycles
and the test of the Adelmans,” Journal of Monetary Economics
33(2), April 1989, pp. 405–438.
462
24 James E.Hartley, Kevin D.Salyer and Steven M.Sheffrin,
“Calibration and real business cycle models: An unorthodox
experiment,” Journal of Macroeconomics 19(1), Winter 1997, pp.
1–17.
496
25 Martin Eichenbaum, “Real business-cycle theory: Wisdom or
whimsy,” Journal of Economic Dynamics and Control 15(4), October
1991, 607–626.
513
26 Gary D.Hansen and Edward C.Prescott. “Did technology
shocks cause the 1990–1991 recession?” American Economic
Review 83(2), May 1993, pp. 280–286.
533
Part VII The Solow residual
27 Robert M.Solow, “Technical change and the aggregate
production function,” Review of Economics and Statistics 39(3),
August 1957, pp. 312–320.
543
28 N.Gregory Mankiw, “Real business cycles: A new Keynesian
perspective,” Journal of Economic Perspectives 3(3), Summer
1989, pp. 79–90.
552
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
x
CONTENTS
29 Zvi Griliches, “The discovery of the residual: A historical
note,” Journal of Economic Literature 34(3), September 1996,
pp. 1324–1330.
564
30 Timothy Cogley and James M.Nason, “Output dynamics
in real business cycle models,” American Economic Review
85(3), June 1995, pp. 492–511.
571
Part VIII Filtering and detrending
31 Robert J.Hodrick and Edward C.Prescott, “Postwar US
business cycles: An empirical investigation,” Journal of Money,
Credit and Banking 29(1), February 1997, pp. 1–16.
593
32 A.C.Harvey and A.Jaeger. “Detrending, stylized facts and the
business cycle,” Journal of Applied Econometrics 8(3), 1993,
pp. 231–247.
609
33 Timothy Cogley and James M.Nason. “Effects of the
Hodrick-Prescott filter on trend and difference stationary
time series: Implications for business cycle research,” Journal
of Economic Dynamics and Control 19(1–2), January–February
1995, pp. 253–278.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
626
Acknowledgements
The editors have benefited from the help of a number of people in putting
this volume together. We thank: John Muellbauer for comments on a
draft of Chapter 1; Michael Campbell for bibliographic work; Jeannine
Henderson for her careful work in preparing camera-ready copy of the
articles reproduced here; Alan Jarvis, Alison Kirk, and Laura Large of
Routledge for their encouragement and help with the production details;
Colin Cameron, Timothy Cogley, Martin Eichenbaum, Robert Feenstra,
Andrew Harvey, Dale Heien, Richard Howitt, Finn Kydland, Martine
Quinzii, Steven Sheffrin, and Gregor Smith for helping us to obtain copies
of the articles suitable for the reproduction in the volume.
We also acknowledge the following permissions to reprint:
The Econometric Society and Dickey and Fuller for a table from “Liklihood
Ratio Statistics for Auto-regression Time Series with a Unit Root” in
Econometrica, 49, 1981, pp. 12–26. The Econometric Society and N.E.
Savin and K.J.White for a table from “The Durbin-Watson Test for Serial
Correlation with Extreme Small Samples or Many Regressors”, Econometrica,
45, 1977, 1989–1986 as corrected by R.W.Farbrother, Econometrica, 48,
September 1980, p. 1554. The American Economic Association and Finn
E.Kydland and Edward C.Prescott for “The Computational Experiment:
An Econometric Tool” in Journal of Economic Perspectives, vol. 10:1, Winter
1996, pp. 69–86. The American Economic Association and Lars Peter
Hansen and James H.Heckman for “The Empirical Foundations of
Calibration” in Journal of Economic Perspectives, vol. 10:1, Winter 1996, pp.
87–104. The American Economic Association and Gary D.Hansen and
Edward C.Prescott for “Did Technology Shocks Cause the 1990–1991
Recession” in American Economic Review, vol. 83:2, May 1993, pp. 280–286.
The American Economic Association and Lawrence J. Christiano and
Martin Eichenbaum for “Current Real-Business Cycle Theories and
Aggregate Labour Market Functions” in American Economic Review, vol.
82:3, June 1992, pp. 430–450. The American Economic Association and
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
XII
ACKNOWLEDGEMENTS
Thomas F.Cooley and Gary D.Hansen for “The Inflation Tax in a Real
Business Cycle Model” in American Economic Review, vol. 79:4, September
1989, pp. 733–748. The Econometric Society and Finn E. Kydland and
Edward C.Prescott for “Time to Build and Aggregate Fluctuations”, in
Econometrica, vol. 50:6, November 1982, pp. 1345–1369. Reed Elsevier
Plc and Edward C.Prescott for “Theory Ahead of Business Cycle
Measurement” in Federal Reserve Bank of Minneapolis Quarterly Review, vol.
10:4, Fall 1986, pp. 9–22. Reed Elsevier Plc and Lawrence H. Summers
for “Some Sceptical Observations On Real Business Cycle Theory” in
Federal Reserve Bank of Minneapolis Quarterly Review, vol. 10:4, Fall 1986, pp.
23–27. Reed Elsevier Plc and Edward C.Prescott for “Response to a Sceptic”
in Federal Reserve Bank of Minneapolis Quarterly Review, vol. 10:4, Fall 1986,
pp. 28–33. Elsevier Science BV, The Netherlands and Robert G.King,
Charles I.Plosser and Sergio T.Rebelo for “Production, Growth and
Business Cycles I: The Basic Neoclassical Model” in Journal of Monetary
Economics, vol. 21:2, March 1988, pp. 195–232. Elsevier Science BV, The
Netherlands and Gary D.Hansen for “Indivisible Labour and the Business”
in Journal of Monetary Economics, vol. 16:3, November 1985, pp. 309–328.
Reed Elsevier Plc and Gary D. Hansen and Randall Wright for “The
Labour Market in Real Business Cycle Theory” in Federal Reserve Bank of
Minneapolis Quarterly Review, Spring 1992, pp. 2–12. Basil Blackwell Ltd
and Finn E.Kydland and Edward C.Prescott for “The Econometrics of
the General Equilibrium Approach to Business Cycles” in Scandinavian
Journal of Economics, vol. 93:2, 1991, pp. 161–178. Oxford University Press
Journals and Kevin D. Hoover for “Facts and Artifacts: Calibration and
the Empirical Assessment of Real-Business Cycle Models” in Oxford Economic
Papers, vol. 47:1, March 1995, pp. 24–44. The American Statistical
Association and Allan W.Gregory and Gregor W.Smith for “Calibration
as Testing: Inference in Simulated Macroeconomic Growth Models” in
Journal of Business and Economic Statistics, vol. 9:3, July 1991, pp. 297–303.
University of Chicago Press Journals and Mark W.Watson for “Measures
of Fit for Calibrated Models” in Journal of Political Economy, vol. 101:6,
December 1993, pp. 1011–1041. John Wiley & Sons Ltd and Fabio Canova
for “Statistical Inference in Calibrated Models” in Journal of Applied
Econometrics, 9, 1994, pp. 123–144. University of Pennsylvania and Fabio
Canova for “Sensitivity Analysis and Model Evaluation in Simulated
Dynamic General Equilibrium Economies” in International Economic Review,
vol. 36:2, May 1995, pp. 477–501. Reed Elsevier Plc and Finn E.Kydland
and Edward C.Prescott for “Business Cycles: Real Facts and a Monetary
Myth” in Federal Reserve Bank of Minneapolis Quarterly Review, vol. 14:2,
Spring 1990, pp. 3–18. University of Pennsylvania and Sumru Altu for
“Time-to-Build and Aggregate Fluctuations”, in International Economic Review,
vol. 30:4, November 1989, pp. 889–920. Fabio Canova, M.Finn, and
A.R.Pagan for “Evaluating a Real Business Cycle Model” © Colin
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
ACKNOWLEDGEMENTS
XIII
P.Hargreaves 1994, reprinted from Nonstationary Time Series Analysis and
Cointegration, edited by C.Hargreaves (1994) by permission of Oxford
University Press. Elsevier Science BV, The Netherlands and Robert
G.King and Charles I.Plosser for “Real Business Cycles and the Test of
the Adelmans” in Journal of Monetary Economics, vol. 33:2, April 1989, pp.
405–438. Louisiana State University Press and James E.Hartley, Kevin
D.Salyer and Steven M.Sheffrin for “Calibration and Real Business Cycle
Models: An Unorthodox Experiment” in Journal of Macroeconomics, vol.
19:1, Winter 1997, pp. 1–17. Elsevier Science BV, The Netherlands and
Martin Eichenbaum for “Real Business Cycle Theory: Wisdom or
Whimsy”, Journal of Economic Dynamics and Control, vol. 15:4, October 1991,
pp. 607–626. MIT Press Journals and Robert M.Solow for “Technical
Change and the Aggregate Production Function” in Review of Economics
and Statistics, 39, August 1957, pp. 312–320. The American Economic
Association and N.Gregory Mankiw for “Real Business Cycles: A New
Keynesian Perspective” in Journal of Economic Perspectives, vol. 3:3, Summer
1989, pp. 79–90. The American Economic Association and Zvi Griliches
for “The Discovery of the Residual: A Historical Note”, Journal of Economic
Literature, vol. 34:3, September 1996, pp. 1330–1334. The American
Economic Association and Timothy Cogley and James M.Nason for
“Output Dynamics in Real Business Cycle Models” in American Economic
Review, vol. 85:3, June 1995, pp. 492–511. Ohio State University Press
and Robert J.Hodrick and Edward C. Prescott “Postwar US Business
Cycles: An Empirical Investigation”, in Journal of Money Credit and Banking,
vol. 29:1, February 1997, pp. 1–16. John Wiley & Sons Ltd and A.C.Harvey
and A.Jaeger for “Detrending, Stylized Facts and the Business Cycle” in
Journal of Applied Econometrics, 8, 1993, pp. 231–247. Elsevier Science BV
and Timothy Cogley and James M.Nason for “Effects of the HodrickPrescott Filter on Trend and Difference Stationary Time Series: Implications
for Business Cycle Research”, Journal of Economic Dynamics and Control, vol.
19:1–2, January–February 1995, pp. 253–278.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
Part I
Introduction
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
Chapter 1
The limits of business cycle research
“That wine is not made in a day has long been recognized by economists.”
With that declaration in Kydland and Prescott’s “Time to Build and
Aggregate Fluctuations” (1982 [3]: 1345),* the real business cycle school
was born. Like wine, a school of thought is not made in a day. Only after
it has matured is it possible to judge whether it is good and to separate the
palatable portions from the dregs. The literature on real business cycle
models has now sufficiently aged, ready for the connoisseurs to pass
judgment.
To facilitate those judgments, we have collected together in this volume
thirty-one previously published articles relevant to real business cycle
models. While there has been no shortage of commentaries on the real
business cycle program, the commentaries have been widely scattered
and have often focused on narrow aspects of the models or represented
partisan positions. Until now, there has not been an easily accessible means
for students of business cycles to assess the real business cycle program on
the basis of the original sources from the perspectives of the critics as well
as the proponents. To date, the most systematic accounts of the real business
cycle program are found in the works of active proponents, particularly in
Thomas Cooley’s (ed.) Frontiers of Business Cycle Research (1995b), and in
the programmatic manifestoes of Kydland and Prescott (1991 [12], 1996
[13]). Yet the critical literature is burgeoning.
The present volume brings together the important articles which make
the case for and against real business cycle models. The articles begin
with the classics of the real business cycle school, starting with Kydland
and Prescott’s (1982 [3]) seminal model. In addition, we include articles
on the methodology of real business cycle models, particular aspects of
the program (e.g., calibration, the measurement of technology shocks,
*
Throughout Chapters 1 and 2, the bold numbers in the square brackets within
the text references refer to later chapters in this volume. However, all page
numbers in these references are the page numbers from the original article.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
4
INTRODUCTION
methods of detrending), as well as articles that attempt to evaluate the
empirical success of the real business cycle model.
The real business cycle program is still a very active one. We therefore
hope that this anthology will prove useful to students and professional
macroeconomists working on real business cycle models—bringing some
perspective to the literature and pointing the way to further research. As
an aid to research, the articles are reprinted here as facsimiles rather than
reset. The preservation of the original pagination, facilitating authoritative
citations, more than compensates, we believe, for the loss of an aesthetically
pleasing typographical consistency.
It is difficult for the neophyte in any area of economics to jump into
the middle of a literature that was meant to advance the current interests
of established economists, rather than a didactic purpose. In the remainder
of this introductory chapter, we aim to provide a segue from the common
knowledge of the advanced student of macroeconomics (or of the
nonspecialist professional) to the essential elements of the real business
cycle program. The objective is to provide a clear, accessible background
to the literature that avoids unnecessary technical complications. At the
same time, in this introductory essay we present our own assessment of
the successes and failures of the real business cycle program. It is an
assessment with which many economists will strongly disagree. We
nevertheless hope that it will be easier for others to articulate their own
assessments against the background of our effort. The articles reprinted
in the volume provide the necessary raw materials.
The technical demands of real business cycle models are often very
high. As a further aid to the neophyte reader of the literature, and to the
potential user of the models, the second introductory chapter to the volume
is a user’s guide to real business cycle models, which provides a step-bystep account of how to formulate, solve, and simulate a real business cycle
model.
So much for preliminaries; let us turn now to the background of the
real business cycle program and to the assessment of its successes and
failures.
I THE REAL BUSINESS CYCLE CONJECTURE
The philosopher of science Karl Popper (1959, 1972) argued that science
progresses through a series of bold conjectures subjected to severe tests.
Most conjectures are false and will be refuted. The truth, by definition,
will survive the ordeal of testing and emerge unrefuted at the end of
inquiry in an infinitely distant future. The boldest conjectures are often
the most fruitful, because, making the strongest claims, they are the most
readily refuted and their refutation narrows the universe of acceptable
conjectures most rapidly. We argue that real business cycle models are
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE LIMITS OF BUSINESS CYCLE RESEARCH
5
bold conjectures in the Popperian mold and that, on the preponderance
of the evidence (to use a legal phrase), they are refuted. It is not, however,
straightforward to see this, because the real business cycle conjecture is
advanced jointly with a claim that models should be assessed using a
novel strategy. We must therefore evaluate the conjecture and the assessment
strategy simultaneously.
Since the publication of Kydland and Prescott’s “Time to Build and
Aggregate Fluctuations” (1982 [3]), the paradigm real business cycle model,
a large and active group of new classical macroeconomists have elaborated
and developed the real business cycle model. As important as these
developments are to the real business cycle program, none of them
fundamentally affects the critical points that we will make.1 Our assessment
will, therefore, focus on the original Kydland and Prescott model and its
successor models in a direct line. We will also refer frequently to the
programmatic statements and methodological reflections of Kydland,
Prescott and Lucas, the most articulate defenders of the aims and methods
of equilibrium business cycle models.
(i) Equilibrium business cycles
To common sense, economic booms are good and slumps are bad. Economists
have attempted to capture common sense in disequilibrium models: full
employment is modeled as an equilibrium: that is, as a situation in which
each worker’s and each producer’s preferences (given his or her constraints)
are satisfied, while anything less than full employment represents a failure
of workers or employers or both to satisfy their preferences. The real business
cycle model is an extraordinarily bold conjecture in that it describes each
stage of the business cycle—the trough as well as the peak—as an equilibrium
(see, for example, Prescott, 1986a [4]: 21). This is not to say that workers and
producers prefer slumps to booms. We all prefer good luck to bad.2 Rather
it is to deny that business cycles represent failures of markets to work in the
most desirable ways. Slumps represent an undesired, undesirable, and
unavoidable shift in the constraints that people face; but, given those
constraints, markets react efficiently and people succeed in achieving the
best outcomes that circumstances permit.
Some other models have come close to the real business cycle conjecture.
Models of coordination failure treat booms and slumps as two equilibria
and posit mechanisms that push the economy to one equilibrium or the
other (e.g., Cooper and John, 1988; Bryant, 1983). Since the boom
equilibrium is the more desirable, policies might seek to affect the
mechanism in a way that improves the chances of ending in the boom
state. The Phillips-curve models of Milton Friedman (1968) and Robert
Lucas (1972, 1973) envisage people achieving their preferences conditional
on an incorrect understanding of the true situation. Booms occur when
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
6
INTRODUCTION
workers believe that real wages are higher than they really are, inducing
them to supply more labor than they would if they knew the truth; slumps
occur when workers believe that real wages are lower than they really are.
Were people fully informed, there would be no booms or slumps.3
The real business cycle model is starker. As with Lucas’s monetary
model, every stage of the business cycle is a Pareto-efficient equilibrium,
but the source of the fluctuations is not misperceptions about prices or the
money supply, but objective changes in aggregate productivity (so-called
technology shocks). Thus, in the midst of a slump (i.e., a bad draw), given
the objective situation and full information, every individual, and the
economy as a whole, would choose to be in a slump.
Contrary to the claims of some proponents of the real business cycle
(e.g., Hodrick and Prescott, 1997 [31]: 1), there is no pre-Keynesian
historical precedent for viewing business cycles as equilibria. Kydland
and Prescott (1991 [12]) see such a precedent in the business cycle models
of Ragnar Frisch (1933), while Lucas (1977:215; 1987:47 inter alia) sees
such a precedent in the work of Hayek (1933, 1935) and other members of
the Austrian School. Hoover (1988, ch. 10; 1995 [15]) demonstrates that
these precedents are, at best, superficial. Frisch’s business cycle models
are aggregative and do not involve individual optimization, even of a
representative agent. Some Austrians reject the notion of equilibrium
altogether. Hayek, who is not among these, accepts dynamic equilibrium
as an ideal case, but sees business cycles as the result of mismatches of
capital type and quantity to the needs of production transmitted to
unemployment through a failure of wages and prices to adjust to clear
markets in the short run—clearly a disequilibrium explanation.4 The real
business cycle model advances a novel conjecture as well as a bold one.
(ii) The novelty of the real business cycle model
Novel in their bold conjecture, real business cycle models nonetheless
have precursors. The primary antecedent is Robert Solow’s (1956, 1970)
neoclassical growth model. In this model, aggregate output (Y) is produced
according to a constant-returns-to-scale production function Φ(•) using
aggregate capital (K), aggregate labor (L), and a production technology
indexed by Z:5
(1.1)
Consumption follows a simple Keynesian consumption function:
(1.2)
where s is the marginal propensity to save. Since Solow was interested in
long-term growth, he ignored the aggregate demand pathologies that
concerned earlier Keynesian economists and assumed that people’s plans
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE LIMITS OF BUSINESS CYCLE RESEARCH
7
were coordinated so that savings (S) equaled investment (I) ex ante as well
as ex post:
(1.3)
Capital depreciates at rate δ and grows with investment:
(1.4)
where indicates the rate of change of capital. Labor grows exogenously
at a rate n per cent per unit time, and labor-augmenting technology (Z)
improves at a rate ␨ percent per unit time, so that effective labor grows at
n+␨.
Under these circumstances, the economy will converge to a steady state
in which the growth of capital after compensating for depreciation is just
enough to match the growth of effective labor. Along the steady-state
growth path both capital and effective labor grow at a rate n+␨; and, since
both inputs to production are growing at that steady rate, so is output
itself.
In the Solow growth model we need to distinguish between equilibrium
and steady state. The model is always in equilibrium, because ex ante
savings always equals ex ante investment (equation (1.3)). But the model
need not be in steady state (i.e., growing at n+␨). Any change in the data
that drives the economy away from the steady state (for example, a change
in s or n) will also produce changes in capital and output (adjustments to
a new steady state), but the economy remains in continuous equilibrium
along the adjustment path.
Lucas (1975) employed the Solow growth model to solve a difficulty in
his own analysis of business cycles. Lucas (1972, 1973) explained the
business cycle as the reaction of workers and producers to expectational
errors induced by monetary policy. The difficulty was to explain why
such expectational mistakes should not be corrected quickly so that business
cycles were short week-to-week, month-to-month, or quarter-to-quarter
fluctuations rather than the five- to six-year cycles typically observed.
Lucas’s solution was to distinguish, in Ragnar Frisch’s (1933) useful
terminology, between impulses that begin a business cycle and propagation
mechanisms that perpetuate a cycle. Expectational errors were the impulses.
These impulses drove the economy away from steady state. Ex post the
economy was seen to be in disequilibrium until the expectational errors
were corrected. But even when they had been corrected, the economy
was returned to an equilibrium away from the steady state. The process of
adjusting capital in order to regain the steady state would be a relatively
slow one. This was the propagation mechanism.
In keeping with the new classical agenda of reducing macroeconomics
to microeconomic foundations, Lucas replaced the stripped-down demand
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
8
INTRODUCTION
behavior of the Solow growth model with the assumption that the behavior
of the aggregate economy can be described by the utility-maximizing
choices of a representative agent, who chooses consumption and labor supply
by solving a dynamic, intertemporal optimization problem. In effect, the
simple consumption function (equation (1.2)) was replaced by a permanentincome (or life-cycle) consumption function, and investment was replaced
by a neoclassical investment function in which the opportunity cost of
capital determines the rate of investment. Unlike in the Solow model, the
factors important to the savings decision now enter separately from those
important to the investment decision. Aggregate demand pathologies are,
nonetheless, impossible, because in Lucas’s model the same agents make
both the savings and the investment decision, which insures ex ante
coordination, and the agents have rational expectations, which insures that
mistakes about the future course of the economy are necessarily
unsystematic. Furthermore, the supply of labor responds elastically to
temporarily high real wages: workers make hay while the sun shines.
Kydland and Prescott’s (1982 [3]) seminal real business cycle model is
a direct outgrowth of Lucas’s monetary growth model. It differs from the
Lucas model in that there is no monetary sector; technology shocks (i.e.,
deviations of Z in equation (1.1) from trend) supply the impulse to business
cycles. The model does not rely on expectational errors. There is no
need. Lucas originally posited expectational errors as a way of permitting
changes in the stock of money to have real effects on the economy without
violating the assumption that money is neutral in the long run. In Kydland
and Prescott’s model, technological change has real effects regardless of
whether it is anticipated. While some of the variability in aggregate output,
consumption, investment, and labor supply in Kydland and Prescott’s
model is attributable to the unexpectedness of technology shocks, the
aggregate variables would fluctuate even if technological change were
perfectly anticipated.
In a recent summary of the real business cycle methodology, Kydland
and Prescott (1997:210) state that “we derive the business cycle implications
of growth theory.” Seen in context, this is misleading. Historically, it is
not the use of the growth model that distinguishes the real business cycle
model from earlier business cycle models. Rather it is finding the impulses
in technology shocks, and modeling the economy in continuous
equilibrium. Both in their theoretical development and (as we shall see
presently) in their quantitative implementation, real business cycle models
abstract from the traditional concerns of growth theory. They provide no
analysis of the steady-state rate of growth at all, but take the factors that
determine it as exogenously given.6 Instead, the focus is on the deviations
from the steady state. Only if growth theory were synonymous with
aggregate general equilibrium models with an optimizing representative
agent would it be fair to say that their behavior is the implication of
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE LIMITS OF BUSINESS CYCLE RESEARCH
9
growth theory. But Solow’s growth model is an example of an aggregate
general equilibrium model that does not posit an optimizing representative
agent. Technology shocks would be propagated in a Solow growth model,
though rather slowly, for the convergence time to steady state is long in a
realistically parameterized Solow model (cf. Sato, 1966). The characteristic
business cycle behavior in real business cycle models comes from the
shocks and from the optimizing model itself (of which more presently),
rather than from the fact that these are embedded in a growth model.
(iii) A quantified idealization
Real business cycle models are implemented by giving specific functional
forms to the equations of the optimal growth model. This is most easily
seen for production; equation (1.1) is replaced by a specific function, very
often the Cobb-Douglas production function:
(1.1´)
where ␪ is the share of labor in national output. An equation such as
(1.1´) could be estimated as it stands or jointly with the other equations in
the model to determine the value of ␪. Real business cycle proponents do
not typically estimate the parameters of their models. Instead, they assign
values to them on the basis of information from sources outside the model
itself. This is known as calibration of the model. The value chosen for ␪ is
usually the average value that the labor share takes in suitably adapted
national accounts.7 The value of the depreciation rate (␦) is calibrated
similarly. As we mentioned already equations (1.2) and (1.3), which
represent aggregate demand in the Solow growth model, are replaced in
real business cycle models by an optimization problem for a representative
agent who is both consumer and producer. The representative agent
maximizes a utility function:
(1.5)
subject to current and future production constraints given by equation
(1.1´), and linked together by equation (1.4), which governs the evolution
of the capital stock. The set {Ct} is the set of current and future levels of
consumption, and {Lt} is the set of current and future supplies of labor
(the time subscript t=0, 1, 2, …, ∞). The utility function must be calibrated
as well. This is usually done with reference to the parameters estimated in
unrelated microeconomic studies.8
The calibrated model is nonlinear. To solve the model its equations are
typically reformulated as linear approximations around the unknown
steady state. This is the technical sense in which real business cycle models
abstract from the concerns of traditional growth theory; for no explanation
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
10
INTRODUCTION
of the steady state is sought—the focus is on (equilibrium) deviations from
the steady state. The solution to the linearized model is a set of linear
equations for output, consumption, labor, and investment of the form:
(1.6–1)
(1.6–2)
(1.6–3)
(1.6–4)
where the lower-case letters are the deviations from steady state of the
logarithms of the analogous upper-case variables. The coefficients γij are
combinations of the calibrated parameters determined by solving the
model.9
The right-hand variables in equations (1.6–1) to (1.6–4) are called state
variables. They summarize the past evolution of the model economy and
are exogenous in the sense that the representative agent takes them as
given data and conditions his choices upon them (z is exogenous and k is
determined from choices made in previous periods). Equations (1.6–1) to
(1.6–4) detail the outcomes of those choices, that is, how the preferences
of the representative agent interact with the constraints he faces, including
the current state of z and k, to determine output, capital, labor, and
investment.
In the original Kydland and Prescott (1982 [3]) model, the technology
shock, z, was modeled as a random process with parameters chosen to
cause the model to mimic the variance of GNP in the US economy. Since
z was artificial, there was no chance of direct period-by-period or historical
comparisons of the modeled time series in equations (1.6–1) to (1.6–4)
with their real-world equivalents. Kydland and Prescott, however, wished
only to compare the covariances among the modeled time series to those
among the actual series, so this did not seem problematic. Nevertheless,
as Lucas (1987:43–45; cf. Prescott, 1986b [6]: 31) noticed, constructing the
predicted output series to mimic actual output does not provide an
independent test of the model.10 Beginning with Prescott (1986a [4]), real
business cycle models have taken a different tack (cf. Kydland and Prescott,
1988). Solow (1957 [27]) attempted to quantify technical change by using
a production function with constant returns to scale (such as equation
(1.6)) to compute Z. Typically, real business cycle models use the CobbDouglas production function (equation (1.1´)) as follows:
(1.7)
This empirical measure of the technology parameter is known as the Solow
residual. When estimated using actual data, the Solow residual, like the
series used to compute it, has a trend (implying ␨⫽0), and so must be
detrended before being used as an input to the real business cycle model.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE LIMITS OF BUSINESS CYCLE RESEARCH
11
Detrended log(Z) is the state-variable z.11
(iv) The limits of idealization
The real business cycle model does not present a descriptively realistic
account of the economic process, but a highly stylized or idealized account.
This is a common feature of many economic models, but real business
cycle practitioners are bold in their conjecture that such models nevertheless
provide useful quantifications of the actual economy. While idealizations
are inevitable in modeling exercises, they do limit the scope of the virtues
one can claim for a model.
In particular, the real business cycle program is part of the larger new
classical macroeconomic research program. Proponents of these models
often promote them as models that provide satisfactory microfoundations
for macroeconomics in a way that Keynesian models conspicuously fail to
do (e.g., Lucas and Sargent, 1979). The claim for providing
microfoundations is largely based on the fact that new classical models in
general, and real business cycle models in particular, model the
representative agent as solving a single dynamic optimization problem on
behalf of all the consumers, workers, and firms in the economy. However,
the claim that representative agent models are innately superior to other
sorts of models is unfounded. There is no a priori reason to accord real
business cycle models a presumption of accuracy because they look like
they are based on microeconomics. Rather, there are several reasons to be
theoretically skeptical of such models.12
Most familiar to economists is the problem of the fallacy of composition,
which Samuelson’s (1948) introductory economics text prominently
addresses. It is difficult to deny that what is true for an individual may
not be true for a group, yet, representative agent models explicitly embody
the fallacy of composition. The central conceptual achievement of political
economy was to analogize from the concerns of Robinson Crusoe—alone
in the world—to those of groups of people meeting each other in markets.
The complexities of economics from Adam Smith’s invisible hand to Arrow
and Debreu’s general equilibrium model and beyond have largely been
generated from the difficulties of coordinating the behavior of millions of
individuals. Some economists have found the source of business cycles
precisely in such coordination problems. By completely eliminating even
the possibility of problems relating to coordination, representative agent
models are inherently incapable of modeling such complexities.
Problems of aggregation are similar to problems arising from the fallacy
of composition. Real business cycle models appear to deal with
disaggregated agents, but, in reality, they are aggregate models in exactly
the same way as the Keynesian models upon which they are meant to
improve. The conditions under which a representative agent could
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
12
INTRODUCTION
legitimately represent the aggregate consequences of, and be deductively
linked to, the behavior individuals are too stringent to be fulfilled:
essentially all agents must be alike in their marginal responses.13 Because
it is impracticable, no one has ever tried to derive the aggregate implications
of 260 million people attempting to solve private optimization problems.
The real business cycle model thus employs the formal mathematics of
microeconomics, but applies it in a theoretically inappropriate circumstance:
it provides the simulacrum of microfoundations, not the genuine article.
It is analogous to modeling the behavior of a gas by a careful analysis of a
single molecule in vacuo, or, of a crowd of people by an analysis of the
actions of a single android. For some issues, such models may work well;
for many others, they will miss the point completely.14
A significant part of the rhetorical argument for using real business
cycle methodology is an appeal to general equilibrium theory. However,
because the models do not reach a microfoundational goal of a separate
objective function for every individual and firm, the models are at best
highly idealized general equilibrium models. Real business cycle theorists
do not appear to be aware of the degree to which this undermines certain
sorts of claims that can be made for their models. The fact that they do
not provide genuine microfoundations essentially removes any prior claims
that real business cycle models are superior to Keynesian or other aggregate
models.
It is not difficult to understand why general equilibrium theory has
such allure for economists in general and macroeconomists in particular.
The theory provides for an extensive model of the economy with individual
consumers maximizing utility and individual firms maximizing profits,
all interacting in competitive markets, and despite all this complexity, it
can be shown that an equilbrium exists. However, knowing that an
equilibrium point exists is all well and fine, but it doesn’t get you very
far. What else can we tell about the economy from the general equilibrium
framework? The answer to that question turned out to be quite depressing;
as Kirman (1989) subtitled a paper about this state of affairs, “The Emperor
Has No Clothes.”
After showing that an equilibrium point existed, people became interested
in the question of whether it could be shown that the equilibrium was
either unique or stable. In order to answer this question, the shape of the
aggregate excess demand curve had to be determined. In a remarkable
series of papers, Sonnenschein (1973, 1974), Mantel (1974, 1976), Debreu
(1974) and Mas-Colell (1977) showed that in an economy in which every
individual has a well-behaved excess demand function, the only restrictions
on the aggregate excess demand function are that it is continuous,
homogenous of degree zero in prices, and satisfies Walras’ Law. Nothing else
can be inferred. Any completely arbitrary function satisfying those three
properties can be an aggregate excess demand function for an economy of
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE LIMITS OF BUSINESS CYCLE RESEARCH
13
well-behaved individuals. Having an economy in which every single agent
obeys standard microeconomic rules of behavior tells us virtually nothing
about the aggregate economy. For example, not even something as basic as
the Weak Axiom of Revealed Preference carries over from the microeconomic
level to the macroeconomic level. (See Shafer and Sonnenschein (1982) for
a complete, technical discussion of this literature and Kirman (1989) or
Ingrao and Israel (1990) for good nontechnical discussions.)
The implication can be stated in two ways. Even if we know that the
microeconomy is well behaved, we know very little about the aggregate
excess demand function. Or, given a completely arbitrary aggregate excess
demand function satisfying the three characteristics above, we can find a
well-behaved microeconomy that generates that aggregate function.
Kirman’s (1992) article in the Journal of Economic Perspectives was largely
centered on showing how these results invalidated the use of a
representative agent model. There is simply no theoretical justification
for assuming that the excess demand function of a representative agent
bears any resemblance to the excess demand function for an aggregate
economy. If we want to justify the notion that macroeconomics needs
microfoundations by pointing to general equilibrium theory, then these
results derived by general equilibrium theorists unambiguously
demonstrate that the representative agent is flawed. Oddly, we seem to be
simultaneously seeing a situation in which macroeconomists point to
general equilibrium theory as a justification for representative agent models
at the same time as general equilibrium theorists are prominently noting
that the representative agent has no home in the theory.
Thus the implicit claim in real business cycle theory that their
representative agent models provide rigorous microfoundations is incorrect.
Starting with first principles, or general equilibrium theory, only, we can
derive all sorts of macroeconomics. Some form of aggregate structure must
be provided.
Beyond this, Kydland and Prescott argue that the models are designed
to capture some features of the economy while ignoring or even distorting
other features. They hold this to be one of their virtues, and argue that
their failure to capture features that they were not designed to model
should not count against them (Kydland and Prescott, 1991 [12]). We
take this claim seriously. It should, nevertheless, be noted that it
undermines the argument that we trust the answers which the models
give us on some dimensions because they have been successful on other
dimensions (Lucas, 1980:272). Kydland and Prescott (1996 [13]: 72) make
exactly this claim with regard to using the Solow growth model to explore
the business cycle. However, if the dimensions on which we need answers
are ones on which, because of their idealized natures, the models are
false, the success on other dimensions is irrelevant. As a point of logic,
rigorous deductions are useful only if they start with true premises.15
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
14
INTRODUCTION
Idealized models are useful because they are tractable, but only if they
remain true in the features relevant to the problem at hand. Kydland and
Prescott want idealized real business cycle models to provide quantitative
conclusions about the economy. There is nothing in their construction
that insures that they will succeed in doing so.
Thus, part of the boldness of the real business cycle conjecture is the
seriousness with which it takes the idealization of a representative agent.
Although economists, at least since Alfred Marshall, have sometimes used
representative agents as a modeling tool, new classical (and real business
cycle) models expect the representative agent to deliver far more than
earlier economists thought possible. For example, Friedman’s (1957)
explication of the permanent-income hypothesis begins with something
that looks like a representative agent, but Friedman uses the agent only as
a means of thinking through what sorts of variables belong in the aggregate
consumption function. He makes no attempt to derive an aggregate
consumption function from his agent; in fact, he takes pains to note how
different the aggregate function will look from the individual’s function.
Real business cycle models, on the other hand, take the functions of
the representative agent far more seriously, arguing that “we deduce the
quantitative implications of theory for business cycle fluctuations” (Kydland
and Prescott, 1997:211). However, for the reasons described above, these
deductions are not the rigorous working out of microeconomic principles
combined with a serious analysis of heterogeneity and aggregation.
There is nothing in the construction of real business cycle models
which insures that they will succeed in providing accurate quantitative
conclusions. There is nothing that guarantees a priori their superiority.
The proof of the pudding is in the eating: the real business cycle model
must be tested and evaluated empirically.
II TESTING
(i) What are the facts about business cycles?
Before real business cycle models can be tested, we must know precisely
what they are meant to explain. Following Prescott (1986a [4]), advocates of
real business cycle models have redefined the explanandum of business
cycles. As observed already, common sense and the traditional usage of
most economists holds that recessions are periods in which the economy is
suboptimally below its potential. Business cycle theory has thus traditionally
tried to explain what causes output to fall and then rise again. To be sure,
this is not just a matter of output declining: when output declines, one
expects employment, income, and trade to decline as well, and these declines
to be widespread across different sectors.16 Nevertheless, the central fact to
be explained was believed to be the decline and the subsequent recovery,
and not the comovements of aggregate time series.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE LIMITS OF BUSINESS CYCLE RESEARCH
15
Even before the first real business cycle models, new classical
macroeconomics shifted the focus to the comovements. Sargent (1979:256)
offers one definition: “the business cycle is the phenomenon of a number
of important economic aggregates (such as GNP, unemployment and
layoffs) being characterized by high pairwise covariances at the low business
cycle frequencies [two- to eight-year range]…. This definition captures
the notion of the business cycle as being a condition symptomizing the
common movements of a set of aggregates.” Lucas (1977:217) argues that
the movements of any single economic aggregate are irregular, and “[t]hose
regularities which are observed are in the comovements among different
aggregative time series.”17
Real business cycle models view the business cycle in precisely the
same way as Sargent and Lucas. The things to be explained are the
correlations between time series, and the typical assessment of the success
or failure of a model is to compare the correlations of the actual time series
to those that result from simulating the model using artificially generated
series for the technology shock (Z). Formal statistical measures of the
closeness of the model data to the actual data are eschewed. Prescott (1986a
[4]), for example, takes the fact that the model approximates much of the
behavior of the actual aggregates as an indicator of its success. In the case
in which the model data predict an empirical elasticity of output to labor
greater than the theory, Prescott (1986a [4]: 21) argues “[a]n important
part of this deviation could very well disappear if the economic variables
were measured more in conformity with theory. That is why I argue that
theory is now ahead of business cycle measurement.”
Kydland and Prescott (1990 [20]) make similar arguments in opposing
“business cycle facts” to “monetary myths.” For example, the real business
cycle model predicts that the real wage is procyclical: a positive technology
shock (an upward shift of the production function), which is a positive
impulse to output, increases the marginal product of labor, and workers
are paid their marginal products. In contrast, monetary business cycle
models (Keynesian and monetarist) predict countercyclical real wages: a
monetary shock, which is a positive impulse to output, increases aggregate
demand and therefore the demand for labor, which requires a lower
marginal product of labor (a movement along the production function).
Kydland and Prescott (1990 [20]: 13–14) argue that a correlation of 0.35
between money lagged one period and current output is too low to support
the view that money leads output; while a correlation of 0.35 between the
real wage and output is high enough to support the view that the real
wage is procyclical. They argue that if measurements were made in closer
conformity to theory, the second correlation would be higher.18 But, even
as it stands, they take the “business cycle facts” as supporting the real
business cycle model.
Theory may be ahead of measurement. It is well understood that as
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
16
INTRODUCTION
science progresses, theoretical advances improve measurements.
Thermometers were originally atheoretical devices that rested on some
simple untested assumptions, such as the linearity of the relationship
between temperature and the expansion of materials. In time, as the theory
of heat developed, better thermometers were possible because theoretical
understanding permitted corrections for departures from the initial
assumptions and new methods of measurement that, among other things,
permitted the range of temperature measurements to be increased to near
absolute zero, at one end, and to millions of degrees, at the other.
Given the best data in the world, however, simply mimicking the data
is a weak test. Indeed, logically it is fallacious to argue that theory A
implies that data behave as B, data in fact behave as B, therefore A is true.
This is the fallacy of affirming the consequent. It is a fallacy because there is
nothing to rule out incompatible theories C, D, and E also implying B.
Popper’s concentration on refutations is a reaction to this fallacy in the
form in which it was exposed by David Hume (1739): there is no logic
that allows one to move inductively from particular instances to a general
rule; there is no inductive logic analogous to deductive logic. It is a
correct inference that theory A implies data behavior B, data fail to behave
as B, therefore A is false. At best, the data limit the class of theories that
are acceptable. One learns very little from knowing that a theory mimics
the data—especially if it was designed to mimic the data. One needs to
know that the data cannot be mimicked by rival theories. Although real
business cycle models are often shown (without any formal metric) to
mimic actual data, they have rarely been tested against rivals.19
It is usually regarded as a more stringent test of a model that it performs
well on a set of data different from the one used in its formulation. Most
often this means that models are formulated on one sample and then
tested against a completely different sample. Kydland and Prescott
(1997:210) offer a different argument: real business cycle models are
formulated using the “stylized facts” of long-run growth theory and are then
tested, not against a completely different data set, but for their ability to
mimic the short-run business cycle behavior of the same data. While there is
clearly merit in deriving empirically supported implications of one set of
facts for another, this particular test provides very weak support for the
real business cycle model. Many models that are fundamentally different
from the real business cycle model, in that they posit neither continuous
equilibrium nor impulses arising from technology shocks, are consistent
with the “stylized facts” of growth (e.g., the constancy of the labor share
in national income or the constancy of the capital-output ratio).20
(ii) Do real business cycle models fit the facts?
Although it is a weak test to check whether models mimic the facts, it is a
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE LIMITS OF BUSINESS CYCLE RESEARCH
17
useful starting point. The fact that real business cycle models are idealized
presents some difficulties in judging them even on such a standard. As
Kydland and Prescott (1991 [12]: 169) stress, the real business cycle model
is unrealistic in the sense that it aims only to capture certain features of
the data rather than to provide a complete explanation. The econometric
ideal is to provide predictions of dependent variables on the basis of all of
the relevant independent variables, so that whatever errors are left are
truly random and independent of any omitted variables. In contrast, real
business cycle models are driven by a single variable, the technology
shock, and aim to explain the relationships among a number of series (as
in equations (1.6–1) to (1.6–4)) on the basis of this single shock. The
success of the model is to be judged, in Kydland and Prescott’s view, on
its ability to capture selected correlations in the actual data. There is no
claim that it will do well in explaining correlations it was not designed to
capture; nor is there any claim that its errors will be truly random, either
in being mean zero and symmetrically (e.g., normally) distributed or in
being independent from omitted variables.
The dilemma is this: Theories are interpretable, but too simple to match
all features of the data; rich econometric specifications are able to fit the
data, but cannot be interpreted easily. The coefficients of a statistically
well-specified econometric equation indicate the effects on the dependent
variable ceteris paribus of a change in the independent variables. In general,
these effects depend in a complicated way on the parameters of the deep
relations that connect the variables together and generate the observed
data. Lucas (1976) in his famous “critique” of policy analysis noticed the
lack of autonomy of econometrically estimated coefficients and argues, in
particular, that the values of the coefficients would not remain stable in
the face of changes in monetary and fiscal policy regimes.
One solution to the Lucas critique might be to identify the complex
structure of the estimated coefficients. Hansen and Sargent (1980) map
out a strategy for doing this. Essentially, the model is taken to be true and
used, in the classic manner of textbook econometric identification, to
disentangle the “deep” parameters (i.e., the parameters of the theory) from
the estimated coefficients.21 The central difficulty with this strategy as a
means of providing support for real business cycle models is that it does
not work. In the case in which the model imposes more relationships
among the parameters than there are parameters to identify, the model is
said to be overidentified. Statistical tests can be used to assess whether
these redundant relationships can be rejected empirically. Altu (1989
[21]) estimated an econometric version of the real business cycle model
and tested its overidentifying restrictions. They were clearly rejected. This
should not be surprising. An idealized model abstracts from too many of
the features of the world for the resulting specification to meet the
econometric ideal. Not only is it likely that the errors will not show
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
18
INTRODUCTION
irreducible randomness and the appropriate symmetry, but they are
unlikely to be independent of omitted variables. One may, of course, add
additional variables into the regression equations. An econometric
specification with many free parameters (i.e., many independent variables)
will invariably fit better than a calibrated business cycle model. But then
one loses the mapping of the theory onto the estimated coefficients that
helped to disentangle the deep parameters.
Kydland and Prescott advocate a second solution: eschew econometric
estimation altogether. They believe that the advantage of the calibrated
model is that it refers to theoretically interpretable parameters, so that
counterfactual experiments can be given precise meanings: for example,
the effects of a change in the persistence of the technology shock or in the
relative risk aversion of consumers or, in richer real business cycle models,
of government-policy rules (e.g., tax rates) have precise analogues in the
calibrated model. A good model, in Kydland and Prescott’s view, is
unrealistic, in the sense that it will not fit the data in the manner of a
statistically well-specified econometric model, but it will fit with respect
to certain features of interest. Calibration and model structure are adjusted
until the models do well against those features of the data that are of
interest.
The development of the labor market in early real business cycle models
provides an illustration of the strategy. Table 1.1 reproduces from Hansen
(1985 [8]) some statistics for actual data and data generated from simulating
two real business cycle models. Model I is a simple model similar to
Kydland and Prescott (1982 [3]) in which labor is supplied in continuously
variable amounts. The standard deviations of hours worked and
productivity are nearly equal in Model I; while, in the actual data, hours
worked are over 50 percent more variable. Model II is a modification of
Model I in which, to capture the fact that workers typically must either
work a full day or not work, labor must be supplied in indivisible eighthour units. Model II was created in part as an attempt to add realism to
capture a feature that was not well described in Model I. In fact, it succeeds
rather too well: hours are nearly three times as variable as productivity in
Model II. Further developments of the real business cycle model (see,
e.g., Hansen and Wright, 1992 [9]) aim in part to refine the ability to
mimic the data on this point.
A serious case can be made for choosing Kydland and Prescott’s strategy
for dealing with the Lucas critique and favoring idealized models at the
expense of achieving the econometric ideal of complete description of the
data (see Hoover, 1995 [15]). The gain is that one preserves theoretical
interpretability—though only at the cost of a limited understanding of the
actual economy. Real business cycle modelers might respond that the
choice is between limited understanding and no genuine understanding
at all. But this would be too glib. There are at least two barriers to declaring
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
Table 1.1. Summary statistics for actual US data and for two real business cycle models
Source: Hansen (1985 [8]), table 1; standard deviations in parentheses.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
20
INTRODUCTION
the triumph of the real business cycle approach on the basis of the
methodological virtues of idealization.
First, most of the assessments of the success or failure of real business
cycle models have been made in the casual manner exemplified by our
previous discussion of Hansen’s (1985 [8]) divisible and indivisible labor
models, using data no more precise than that of Table 1.1. The standard
is what might be called “aesthetic R2”: whether Models I or II in Table
1.1 are too far from the actual data or close enough is a purely subjective
judgment without a good metric.
One response might be that no formal metric is possible, but that a
more rigorous subjective evaluation would go some way to providing the
missing standards. King and Plosser (1989 [23]) take this tack. They revive
the method of Adelman and Adelman (1959), first used to evaluate the
Klein-Goldberger econometric macromodel. King and Plosser simulate
data from a real business cycle model and evaluate it using the business
cycle dating procedures developed by Burns and Mitchell at the National
Bureau of Economic Research. These techniques aim to characterize the
repetitive features of the economy by averaging over historical business
cycles normalized to a notional cycle length.22 Both the actual data and
the simulated data from the real business cycle model are processed using
Burns and Mitchell’s procedures. King and Plosser observe that it is
difficult to discriminate between these two sets of data. But they note that
the results “leave us uncomfortable,” because the same claims can be made
on behalf of the Keynesian Klein-Goldberger model. Despite the greater
detail in this study compared to typical assessments of real business cycle
models, it is still wedded to aesthetic R2.
In a similar vein, Hartley, Salyer, and Sheffrin (1997 [24]) examine the
ability of the standard informal methods of assessment of real business
cycle models to discriminate between alternative accounts of the actual
economy. Hartley et al. use the Fair macroeconometric model of the US
economy, a model in the tradition of Keynesian macroeconomic forecasting
models such as the Brookings model or the Federal Reserve-University of
Pennsylvania-Social Science Research Council model, to simulate data
for a “Keynesian” economy in which demand shocks and disequilibrium
are important.23 Calibrating a real business cycle to be consistent with the
relevant parameters of the Fair model, they ask whether the real business
cycle model, which is driven by technology shocks (these are calculated
from the simulated data from the Fair model) and continuous equilibrium,
can mimic a “Keynesian” economy. They find out that it can to at least as
high a degree as it mimics the actual economy on the usual standards
used by real business cycle modelers. One interpretation of this result is
that it is very bad news for the real business cycle model, because it shows
that it has no power of discrimination; its key assumptions do not restrict
the sort of economies it can fit.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE LIMITS OF BUSINESS CYCLE RESEARCH
21
A real business cycle modeler, however, might riposte that the Fair
model is a typical Keynesian model with many free parameters, so that it
gives a good statistical description of the economy, even as it fails to
model the true underlying mechanisms. Thus the fact that the real business
cycle model “works” for simulations from the Fair model means nothing
more than that it works for the actual economy. To check this interpretation,
Hartley et al. alter two key parameters, those governing the interest
elasticities of money demand and investment, changes which makes the
simulations of the Fair model (particularly, the cross-correlations stressed
by real business cycle analysts) behave more like European economies
than like the US economy (see Backus and Kehoe, 1992). The real business
cycle model is poor at mimicking the data from the altered Fair model.
One might conclude that the real business cycle model is, in fact,
discriminating. However, for a modeling strategy that takes pride in its
grounding in fundamental and universal economic theory (the Solow
growth model is not country-specific), this is hardly an attractive
conclusion. Although European business cycles may be substantially
different from American business cycles because of important institutional
differences, real business cycle models typically seek to explain business
cycles abstracting from those very institutional details.
A second barrier to declaring the triumph of the real business cycle
model on the basis of the methodological virtues of idealization is that,
even if idealized models cannot be expected to fit as well as traditional
econometric specifications under the best of circumstances, the conclusion
that econometric estimation is irrelevant to the real business cycle model
would be unwarranted. Calibration might be regarded as a form of
estimation (Gregory and Smith, 1990; 1991 [16]). The problem is how to
judge the performance of calibrated models against an empirical standard.
Watson (1993 [17]) develops an asymmetrical measure of goodness of fit
that is useful for real business cycle models precisely because their idealized
nature makes it likely that the errors in fitting them to actual data are
systematic rather than random. Even using his goodness of fit measure,
Watson fails to produce evidence of high explanatory power for real
business cycle models.
Kydland and Prescott’s (1991 [12]) objection to traditional econometrics,
what they call the “systems of equations” approach, is that an idealized
model will not provide the necessary restrictions to permit the accurate
estimation of its own parameters on actual data, because of the many
features of the data that it systematically and intentionally ignores. Canova,
Finn and Pagan (1994 [22]) undertake a somewhat less demanding test.
Where Altug (1989 [21]) had tested restrictions that were strong enough
to identify all the parameters of the real business cycle model and, therefore,
to eliminate the need for calibration, Canova et al. examine the implications
of a previously calibrated real business cycle model for the dynamic
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
22
INTRODUCTION
behavior of various time series. They observe that the various time series
can be described by a vector autoregression (VAR) in which each series is
regressed on its own lagged values and the lagged values of each of the
other series. What is more, if the real business cycle model is an accurate
description of the actual data, then a number of restrictions must hold
among the estimated parameters of the VAR.
The real business cycle model implies three sets of restrictions on the
VAR of two distinct types. First, various time series should be cointegrated.
Two series are cointegrated when a particular linear combination of them
is stationary (i.e., when its mean, variance, and higher moments are
constant), even though the series are not separately stationary. There are
two sets of such cointegration restrictions: (1) the state variables (the
analogues of Z and K, the non-detrended counterparts to the state variables
in equations (1.6–1) to (1.6–4)) must stand in particular linear relationships;
(2) state variables and predicted values for various time series (e.g., the
left-hand variables in equations (1.6–1) to (1.6–4)) to must also stand in
particular linear relationships. Finally, once one has accounted for the
cointegrating relationships among these time series and concentrates on
their behavior about their common trends, there is a third set of restrictions
(second type), which are the particular implications of the real business
cycle model for the parameters of the VAR.
Canova et al. use a calibrated real business cycle model with a
considerably richer specification than Kydland and Prescott’s early models
to derive the necessary restrictions on the VAR. These restrictions are
then compared to the data. Canova et al. show that the restrictions do not
hold. A particularly interesting finding is that the real business cycle
model imputes too much importance to the productivity shock.
Canova et al.’s imposition of a specific numerical calibration of the real
business cycle model might limit the generality of their results: it might
be said that the real business cycle model is correct in principle, but
Canova et al. have failed to calibrate it correctly. In defense of their test,
their choice of parameters is not at all atypical. What is more, they examine
a limited range of alternative choices of parameters, asking the question:
What parameters would it take to make the model agree with the data?
Their results along these lines, however, are not nearly as comprehensive
as they would need to be to close the case.
Eichenbaum (1991 [25]) examines the issue of parameter choice more
systematically. He begins by noting that the numerical values of the
underlying parameters used to calibrate a real business cycle model are
simply estimates of the true values. We do not know the true values of
things like the depreciation rate or the variance of the shock to the Solow
residual. Instead, we estimate these numbers from sample data, and there
is a sampling error associated with every estimate. (Hansen and Heckman,
1996 [14]: 95, make a similar argument.)
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE LIMITS OF BUSINESS CYCLE RESEARCH
23
Eichenbaum finds that altering most of the parameters within the range
of their sampling error does little to alter the behavior of the real business
cycle model. The notable exceptions are the parameters associated with
the Solow residual, which have large standard errors. He finds that at
standard levels of statistical significance (5 percent critical values),
technology shocks may account for as little as 5 percent and as much as
200 percent of the variance in output. Eichenbaum’s results suggest that,
even if real business cycle models had no other problems, we cannot
reject the view that technology shocks in conjunction with a real business
cycle model explain only a small fraction of aggregate fluctuations.
Although not decisive, conventional econometric tests of real business
cycle models are not kind to the theory. Canova et al.’s investigation of
alternative real business cycle specifications, like Hartley et al.’s investigation
of alternative data-generating processes, reminds us that no test of the real
business cycle is likely on its own to provide a decisive Popperian refutation.
The very fact that the models are idealized implies that the actual data
alone provide at best a weak standard. More important than simply fitting
the data is the relative performance of alternative models. Canova et al.
and Hartley et al. push in the right direction, though not terribly far. Of
course, the advocates of real business cycle models have always judged
them relatively against other models in their class. Hansen’s (1985 [8])
model with indivisible labor was judged superior to his model with
divisible labor. Cooley and Hansen (1989 [11]) present a real business
cycle model with a monetary sector and additional monetary shocks;
Christiano and Eichenbaum (1992 [10]) present one with a government
sector and fiscal policy shocks. Other models have included household
production (Benhabib, Rogerson, and Wright, 1991) or variable capacity
utilization (Greenwood, Hercowitz, and Huffman, 1988).
All of these models, however, retain the common core of the original
Kydland and Prescott real business cycle model. The only substantial
comparison between a real business cycle model and one with quite
different principles of construction is found in Farmer’s (1993) model of
an economy with increasing returns to scale and shocks to “animal spirits.”
In Farmer’s model there are multiple equilibria. The economy ends up in
one equilibrium or another depending upon the self-fulfilling expectations
of consumers. Farmer argues that his model performs better than the real
business cycle model using Kydland and Prescott’s standard of mimicking
the relative correlations of actual data. He also claims that his model
captures the dynamics of the economy more accurately. He estimates vector
autoregressions for the actual economy and then uses the estimated
equations to generate the path the economy would follow as the result of
shocks to the various variables (i.e., impulse response functions). He then
compares the impulse response functions of the real business cycle model
and of his model with multiple equilibria to each other and to those of the
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
24
INTRODUCTION
estimated VARs. He finds that the impulse responses of the real business
cycle model are very different from his model and that his model is more
like those from the VAR. Once again, the appeal is to aesthetic R2. Further
work on standards of comparison is much to be desired.24
(iii) Testing the elements of the real business cycle model: the impulse
mechanism
Rather than assessing the performance of the real business cycle model
directly against the data, we can ask how well its fundamental components
succeed. As we noted earlier, one of two distinguishing features of the
real business cycle model is that it locates the impulse to business cycles
in technology shocks. The overarching question is: What evidence do
we have that technology shocks are the principal impulse driving the
business cycle?
Before we can answer that question, however, another more basic one
must be answered: What are technology shocks? This is a question that
has plagued real business cycle research from the beginning (see, e.g.,
Summers, 1986 [5]). At the formal level, technology shocks are just the
deviations of the parameter Z in the aggregate production function (e.g.,
equations (1.1) or (1.1') above) from its steady-state growth path: we
represented these shocks earlier as z. By analogy to the microeconomic
production function for individual products, one might naturally interpret
z as a measure of physical technique or organizational ability.
An aggregate measure should average out shocks to particular technology,
so that z should measure shocks that have widespread effects across the
economy. Such averaging should reduce the variability of the aggregate
shocks relative to the underlying shocks to individual technologies. However,
in order to make the real business cycle model match the variability of US
output, the technology shocks must be relatively large and persistent: Kydland
and Prescott (1982 [3]) model z as an autoregressive process with a half-life
of about 14 quarters and a standard deviation of 2.9 percent of trend real per
capita GDP. Our calculations for the period 1960:1 to 1993:1 are similar,
yielding a standard deviation for z of 2.8 percent and for GDP per capita of
4.4 percent about trend.
Although technology is improving over time, Kydland and Prescott’s
assumptions about the variability of z imply that technology must
sometimes regress. But as Calomiris and Hanes (1995:369–70) write:
“Technological regress does not appear to correspond to any event in
Western economic history since the fall of the Roman Empire.” Elsewhere
they point to the large literature on the introduction and diffusion of
particularly important technologies through history: even for such crucial
technological developments as the steam engine, the electric motor, and
the railroad, the speed of diffusion is relatively slow, so that new
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE LIMITS OF BUSINESS CYCLE RESEARCH
25
technologies take decades rather than quarters to spread through the
economy. Calomiris and Hanes (1995:369):
conclude that the diffusion of any one technological innovation
could not increase aggregate productivity by more than a trivial
amount from one year to the next. If no single innovation can make
much of a difference, it seems extremely unlikely that the aggregate
rate of improvement could vary exogenously over cyclical frequencies
to an important degree.
In the face of such objections, proponents of real business cycle models
have broadened the scope of technology to include “changes in the legal
and regulatory system within a country” (Hansen and Prescott, 1993 [26]:
281). Fair enough; such changes may be important to the economy and
may plausibly be negative; but are they likely to justify quarter-to-quarter
variation in z of the required amount? Furthermore, as Calomiris and
Hanes (1995:370) point out, regulatory and legal intervention in the US
economy was substantially smaller before World War I when business
cycles themselves were more variable.25
Debates over the size and frequency of technology shocks are difficult
to resolve because the shocks are not directly observable. Real business
cycle models have largely adopted the biblical criterion “by their fruits ye
shall know them” and used the Solow residual (equation (1.7) above) as a
proxy for technology shocks. The Solow residual attributes to technology
any change in output that cannot be explained by changes in factor inputs.
Jorgenson and Griliches (1967) and Griliches (1996 [29]) point out that
the Solow residual measures more than underlying technological change
(a fact recognized by Solow, 1957 [27]: 312, himself), picking up among
other things variability in capital utilization and labor hoarding.26
Summers (1986 [5]) and Mankiw (1989 [28]) reiterate these points in the
context of real business cycle models. Hall (1986, 1990) notes that
calibrating the parameters of the Cobb-Douglas production function
(equation (1.1')), ␪ and (1-␪), as the shares of labor and capital in output
in the calculation of the Solow residual (as in equation (1.7)) requires the
assumption of perfect competition so that firms and workers are paid their
marginal products and factor shares exactly exhaust output. But if firms
have market power so that price exceeds marginal cost, factor shares will
no longer coincide with the technological parameters ␪ and (1-␪), and z
will reflect variations in markups across the business cycle as well as true
technology shocks. Hall (1990) also demonstrates that if there are increasing
returns to scale, the Solow residual will move with things other than pure
technology shocks.
Jorgenson, Griliches and Hall conclude that the Solow residual captures
a great deal besides technology. Hartley (1994) provides evidence that the
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
26
INTRODUCTION
Solow residual may not reliably capture even genuine technology shocks.
The evidence is found in simulated economies constructed using Hansen
and Sargent’s (1990) flexible, dynamic linear-quadratic equilibrium
macromodel. This model permits a richer specification of the underlying
production technology than typical of real business cycle models: there
are multiple sectors, including intermediate and final goods, and parameters
representing multiple aspects of the production process. Hartley was able
to generate series for output, capital, and labor based on shocks to specific
parts of the production process. Because these were simulations, he could
be assured that the variability in these series reflected only technology
shocks and not market power, labor hoarding, and the like. He then
calculated Solow residuals from the simulated series using equation (1.7)
and asked whether these accurately reflected the size and direction of the
underlying true technology shocks. For a wide range of plausible
parameters, Hartley found an extremely low correlation between his
controlled technology shocks and the calculated Solow residuals. Often
the correlation was not even positive. The failure of the Solow residual to
capture the underlying process accurately appears to reflect the fact that
the Cobb-Douglas production function, implicit in the calculation of Solow
residuals, is a poor approximation to the rich production details of Hansen
and Sargent’s model: the Solow residuals largely reflect specification error,
rather than technological change on a quarter-by-quarter horizon. Hansen
and Sargent’s model is rich relative to the typical idealized real business
cycle model, but is itself an extreme idealization of the real production
process. Hartley’s simulation results, a fortiori, call the Solow residual
into question as a measure of actual technology shocks.
(iv) Testing the elements of the real business cycle model: the
propagation mechanism
The propagation mechanism of a business cycle model should transmit
and amplify the impulses to the various cyclical aggregates. Together with
the shocks themselves it should account for the pattern of fluctuations in
each series and for their comovements. Real output is generally taken as
the marker series for the business cycle. The balance of evidence is that
real business cycle models add relatively little to the pattern of fluctuations
in real output beyond what is implicit in the technology shocks themselves.
Watson (1993 [17]) uses spectral analysis to decompose the power of the
real business cycle model to match movements in output at different
frequencies or (equivalently) time horizons. He finds that the spectral
power of the real business cycle model is high at low frequencies
(corresponding to trend or long-term growth behavior), but low at business
cycle frequencies (approximately two to eight years). Cogley and Nason
(1995b [30]) compare the dynamic pattern of the technology shocks fed
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE LIMITS OF BUSINESS CYCLE RESEARCH
27
into the real business cycle model with the predicted time series for output
generated by the model. Again, they find that it is the dynamic properties
of the exogenous inputs that determine the properties of the output, with
the model itself contributing almost nothing. In one sense, these results
should not be surprising: the Solow growth model, the foundational model
of the real business cycle model, was originally meant to capture secular
change. It is bold to conjecture that, unaltered, it would also model the
business cycle. What is more surprising is that it took relatively long to
document its low value-added with respect to business cycles.
Part of the reason that the real business cycle model has appeared to do
well is that its proponents—sometimes for good methodological reasons—
have relied on standards of assessment that are not particularly
discriminating and have failed to develop more discriminating ones (see
section II (ii) above). Part of the reason has to do with the standard practices
of real business cycle models with respect to handling data. The real
business cycle model predicts values for output, consumption, investment,
and other time series expressed as deviations from the steady state. In
order to compare these with actual data, an estimate of the steady state
must be removed from these variables, which typically are trending. The
Solow growth model suggests that all these variables should grow at rates
related to the steady-state growth rate. Unfortunately, that is not observable.
In practice, real business cycle models follow one of two strategies to
generate detrended data. They sometimes remove a constant exponential
trend, which is linear in the logarithm of the series, and so is known as
linear detrending (e.g., King, Plosser and Rebelo, 1988 [7]). This would
be precisely correct if the growth model were in fact true and the rate of
growth of the labor force (n) and of technology (␨) were constant over
time, so that the steady-state growth rate (n+␨) were also constant over
time. But there is no reason to think that this is so. An alternative strategy
is to use a slowly varying trend that effectively allows the steady-state
growth rate to be variable. This is the most popular option and it is typically
implemented using the Hodrick-Prescott (HP) filter (Hodrick and Prescott,
1997 [31]).27 The HP filter is a nonlinear regression technique that acts
like a two-side moving average. As we noted, and as Prescott (1986a [4])
asserts, one should in principle model growth and cycles jointly (see also
Kydland and Prescott, 1996 [13]). In practice, however, real business cycle
models express the interrelationships of data as deviations from the steady
state. So, in effect, the HP filter provides an atheoretical estimate of the
steady-state growth path.
Harvey and Jaeger (1993 [32]) analyze the usefulness of the HP filter in
accomplishing this task. They compare the cyclical component for output
generated from an HP filter to that from a structural time-series model in
which the trend and the cyclical component are estimated jointly. (This is
closer to what Kydland and Prescott advocate in principle than to what
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
28
INTRODUCTION
they actually practice.) For US GDP, both detrending methods produce
similar cyclical components. Harvey and Jaeger, however, demonstrate
that the H P filter is wildly different from the structural
time-series model for several other countries. This underscores the
previously cited finding of Hartley et al. (1997 [24]) that the real business
cycle model matches US data—at least on the standard of aesthetic R2
typically employed by real business cycle modelers—but not artificial data
of a more “European” character.
Harvey and Jaeger also show that the HP filter and the structural
timeseries model differ substantially when applied to other US time series—
particularly in the case of U.S. prices and the monetary base. Given
Kydland and Prescott’s impassioned attack on the “monetary myths” of
the business cycle, it is obviously critical to know whether the facts about
money and prices are independent of the filtering process. Furthermore,
Harvey and Jaeger demonstrate that in small samples the HP filter can
induce apparent cyclical fluctuations and apparent correlations between
series even when the prefiltered series are independent and serially
uncorrelated. As they point out, these results are in the same spirit as
Slutsky’s and Yule’s analyses of spurious cyclical behavior (Yule (1926),
Slutsky ([1927] 1937); more recently, see Nelson and Kang, 1981). This
phenomenon has been long known if not fully appreciated. Simon Kuznets,
for example, “discovered” long cycles in US data that had first been
transformed through two separate moving averages and first differencing.
It can be shown that purely random data subjected to such transformations
present precisely the same twenty-year cycles that Kuznets reported: they
are nothing but an artifact of the filtering (see Sargent, 1979:249–51). The
analogy between the HP filter and Kuznets’s transformations is close
because the HP filter acts as a type of two-sided moving average.
Cogley and Nason (1995a [33]) reinforce Harvey and Jaeger’s analysis;
they demonstrate that prefiltered data do not generate cycles in a real
business cycle model, while HP filtered data do. Furthermore, when the
input data are highly serially correlated (a correlation coefficient of unity,
or nearly so, between current and lagged values of the variable: i.e., a
unit root or near unit root), the HP filter not only generates spurious
cycles but also strongly increases the correlation among the predicted
values for output, consumption, investment, hours of work, and other
predicted values from the real business cycle model. The model itself—
that is, the propagation mechanism—does little of the work in generating
the cyclical behavior; the HP filter does the lion’s share.
On the one hand, the use of the HP filter calls into question the very
facts of the business cycle. Kydland and Prescott (1990 [20]) document
the intercorrelations among HP filtered time series. These correlations
are held by real business cycle modelers to provide strong prima facie
support for the real business cycle model (Kydland and Prescott’s (1990
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE LIMITS OF BUSINESS CYCLE RESEARCH
29
[20]) subtitle to their paper is “Real Facts and a Monetary Myth”). For
example, they show that the correlation between HP-filtered real GDP
and HP-filtered prices is -0.50, and claim that this contradicts the
prediction of Keynesian models that prices are procyclical. Harvey and
Jaeger (1993 [32]) not only show that the HP filter can induce such
correlations, but they also show that it adds statistical noise, so that a
genuine correlation would, in a sample size of one hundred, have to
exceed 0.40 before we could be sure that it was statistically significant at
the standard 5 percent critical value. If such correlations are really artifacts
of a filtering procedure, with no particular grounding in the economics
of the business cycle, then the support of the “real facts” for the real
business cycle model is substantially weakened.
Prescott (1986a [4]: 10) wrote: “If the business cycle facts were sensitive
to the detrending procedure used, there would be a problem. But the key
facts are not sensitive to the procedure if the trend curve is smooth.” The
weight of evidence since Prescott wrote this suggests that he is incorrect:
the facts are sensitive to the type of filtering that defines the trend.
On the other hand, while there is good reason to find some way to
detrend the technology-shock series used as an input into the real business
cycle model, it is also standard practice to HP filter the predicted series
generated by the real business cycle model before checking their
intercorrelations and comparing them to the HP filtered actual data. Harvey
and Jaeger’s and Cogley and Nason’s analyses suggest that this practice
raises the correlations among these series artificially.
Kydland and Prescott (1996 [13]: 76–77 n) defend the use of the HP
filter against critics who have argued that it induces spurious cycles by
stating that deviations from trends defined by the HP filter “measure
nothing,” but instead are “nothing more than well-defined statistics”; and,
since “business cycle theory treats growth and cycles as being integrated,
not as a sum of two components driven by different factors… talking
about the resulting statistics as imposing spurious cycles makes no sense.”
The logic of Kydland and Prescott’s position escapes us. It is true that real
business cycle theory treats the business cycle as the equilibrium
adjustments of a neoclassical growth model subject to technology shocks.
But, as we have previously noted, the real business cycle model does not,
in practice, model the steady state. The HP filter is an atheoretical method
of extracting it prior to the economic modeling of the deviations from the
steady state. The implicit assumption is that the extracted trend is a good
approximation of the steady state, for which no evidence is offered. This
does not say that the steady state could not be modeled jointly with the
deviations in principle. That it is not actually modeled jointly in practice,
however, means that the objection to the HP filter raised by many critics
remains cogent. The work of Harvey and Jaeger and Cogley and Nason
(see also Canova, 1991), which Kydland and Prescott wish to dismiss,
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
30
INTRODUCTION
demonstrates that the choice of which ad hoc method is used to extract
the balanced-growth path greatly affects the stochastic properties of the
modeled variables and their relationships with the actual data.
One way of reading Watson (1993 [17]) and Cogley and Nason (1995a
[30]) is that, while a model driven by a technology shocks fits output well,
it is the technology shocks not the model which are responsible for that
fact. The picture painted is one of the real business cycle model as a
slightly wobbly transmission belt converting the time-series characteristics
of the technology shocks into the model’s predictions for real output. But
in the end there is a good fit between the model and real output. King
(1995) and Hoover (1997) suggest that if the Solow residual is taken as the
proxy for technology shocks, then this success is an illusion.
Despite having rejected in earlier work the relevance of direct comparisons
to historical data, real business cycle models have recently made precisely
such comparisons.28 Hansen and Prescott (1993 [26]) ask whether technology
shocks can explain the 1990–91 recession in the United States, while Cooley
(1995a) asks whether they can explain the “Volcker recessions” of 1980–82.
In each case, the predictions of a real business cycle model are compared
directly to the historical path of real output.29
Again the standard is one of aesthetic R2, and the pitfalls of this standard
are easily seen Hansen and Prescott’s (1993 [26]: 285) figure 4 (see p. 538
below). Hansen and Prescott cite the fact that the output predicted from
their real business cycle model tracks actual output as favorable evidence
for its explanatory power. In particular, they note that the model catches
the fall in output in 1990–91. But look more closely. Actual GNP peaks
in the first quarter of 1990, while model GNP peaks in the fourth quarter;
actual GNP bottoms out in the first quarter of 1991, while model GNP
bottoms out in the second quarter. Furthermore, the model predicts two
earlier absolute falls in GNP, while, in fact, there are no other recessions
in the data. One of these predicted falls is actually on a larger scale than
the genuine recession of 1990–91: the model shows that GNP peaks in
the first quarter of 1986 and falls 2.3 percent to a trough in the fourth
quarter of 1986, where in reality GNP rose the entire time. The actual
fall in GNP in the 1990–91 recession is only 1.6 percent.
The difficulties of using aesthetic R2 to one side, these graphical measures
or their more statistical counterparts (e.g., see Smith and Zin, 1997) offer
no support for the real business cycle model. To see the difficulty, consider
a simple version of a real business cycle model in which we abstract from
time trends. Initially, let labor be supplied inelastically. Capital is inherited
from the previous period. The Solow residual (zt) can be calculated in loglinear form:
(1.8)
The log-linear version of the production function is given by
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE LIMITS OF BUSINESS CYCLE RESEARCH
31
(1.9)
where the s subscripts refer to variables determined in the model. From
the inelastic supply of labor we know that Ls=Lt. Substituting this fact and
the Solow residual into equation (1.9),
(1.10)
or
(1.10´)
Our model fits perfectly; the correlation between predicted and actual
output is unity. Does anyone believe that this is a demonstration of its
goodness?
Real business cycle models do not fit perfectly, as this little exercise
suggests. The reason is that the inputs to their production function do
not recapitulate the capital and labor measures used to calculate the Solow
residual. In particular, the labor input (Ls) is determined by other features
of the model—in fact, by features that are considered the most characteristic
of real business cycle models, such as intertemporal substitutability of
labor and optimal investment and consumption planning.30 So, it is natural
to relax our assumption of an inelastic labor supply. Then equation (1.10)
becomes
(1.11)
How well predicted output fits actual output is seen to depend on how
well predicted labor fits actual labor. Still, there is an artifactual element
to the correlation between predicted and actual output. Notice that the
share parameter ␪ is not given in nature, but is a modeling choice. If ␪ is
calibrated to be close to zero, then the predicted and actual output are
again nearly perfectly correlated. Now, it might be objected that we know
␪ is not close to zero but close to 0.69 (Hansen and Prescott’s (1993 [26])
assumption). We agree. But information about the true size of ␪ comes
from the calibrator’s supply of exogenous information and has nothing to
do with the fit of the model to historical data or with traditional
econometrics. It underscores Kydland and Prescott’s advocacy of external
sources of information to pin down free parameters. We must not forget
that whether ␪ is zero, 0.69, or one, actual output shows up on the righthand side of equation (1.11) only because we put it there in the construction
of the Solow residual, not because the model generated it by closely
matching the structure of the economy.31
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
32
INTRODUCTION
Of course, it would be a marvelous testament to the success of the
model if the right-hand side of equation (1.11) turned out to be very
nearly log(Yt). That would occur because the model’s predicted labor was
very nearly the actual labor. The true measure of such success is not
found indirectly in the comparison of Ys to Yt, but in the direct comparison
of Ls to Lt.
Even a test based on modeling labor conditional on the Solow residual
is likely to suffer from spurious success at fitting historical labor, since the
Solow residual also contains current labor information by construction.
Truly revealing tests of the success of the real business cycle model at
capturing the true propagation mechanism based on comparisons of the
predictions of the model against historical time series should then
concentrate on those series (e.g., consumption) the current values of which
play no part in the construction of measures of the technology shocks.
To give up the comparison of historical and predicted output (or labor)
is not to give up the comparison of historical and predicted data altogether.
One might ask different questions of the model: for example, if one knew
actual output and capital, what would the model imply that consumption
and labor would have to be? These conditional predictions are measures
of consumption and labor that are uncorrupted by actual labor in their
construction. Historical comparisons on these dimensions would be useful
tests of the model: a close fit would then be a genuine accomplishment of
the real business cycle model, and not an artifact of the construction of
the Solow residual.32 We know of no work to date that has systematically
investigated the propagation mechanism of the real business cycle model
in this manner independently of the Solow residual.
III REFUTATION?
The history of real business cycle models illustrates a fact well known to
philosophers and historians of science: It is rare for a conjecture—however
bold—to be refuted simpliciter on the basis of a single experiment or a single
observation, as in Popper’s ideal case. Accumulated evidence may,
nonetheless, render the intellectual cost of persisting in a particular
conjecture (model or theory) higher than the cost of abandoning or
modifying it. To some extent, it does not appear to be controversial that
the evidence weighs against the real business cycle program narrowly
construed. Even the best-known advocates of real business cycle models
have tended to move away from models of perfectly competitive
representative agents driven by technology shocks only (see n. 1). While
these models are direct descendants of the real business cycle model and
remain in the broader class of equilibrium business cycle models, they
represent an abandonment of the strongest form of the real business cycle
conjecture. The balance of the evidence presented here is that they are
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE LIMITS OF BUSINESS CYCLE RESEARCH
33
right to do so. Although there can be no objection to investigating just
how far these new models can be pushed, there is little in the evidence
with respect to the narrower real business cycle conjecture that would
warrant much optimism about their success.
The case against the real business cycle conjecture has several parts.
First, the propagation mechanism (i.e., the Solow growth model), while it
provides, to a first approximation, a reasonable account of long-term
growth, has virtually no value added with respect to business cycles. The
growth model will transmit fluctuations at business cycle frequencies from
impulses that are already cyclical, but it will not generate them from noncyclical impulses.
The putative impulse mechanism is the fluctuation of technology. In the
model itself this amounts to shifts in a disembodied parameter (Z). The
proponents of real business cycle models have given very little account of
what features of the world might correspond to Z and fluctuate in the way
needed to produce business cycles. Z is an unexplained residual in every
sense of the word: it is whatever it has to be to make the real business cycle
model behave in an appropriate manner, and it cannot be independently
observed. If measured as the Solow residual, “technology” means whatever
bit of output that cannot be accounted for by capital and labor inputs.
Using this residual output as an impulse cannot yield predicted values for
output that provide a logically sound independent comparison between the
model and the actual data on the dimension of output.
While valid comparisons might be made on other dimensions, the
actual evidence in favor of real business cycles is weak in the sense that it
does not provide discriminating tests: alternative models do as good a job
in mimicking the data on the usual aesthetic standards as does the real
business cycle model. Both the facts to be explained and the ability of the
models to match those facts are themselves frequently distorted by the
common data-handling techniques (particularly the HP filter). These data
problems, combined with the fact that the highly idealized nature of the
real business cycle models limits the ambitions that their advocates have
for matching the actual data, insulate the model from decisive refutation,
but equally well undercut the role of empirical evidence in lending positive
support to them.
The real business cycle model has for fifteen years dominated the agenda
of business cycle research. On balance, however, there is little convincing
empirical evidence that favors it over alternative models. To its advocates,
the paucity of evidence may not be of too much concern, for Kydland and
Prescott (1991 [12]: 171) argue that the confidence one places in a model
to answer economic question cannot “be resolved by computing some
measure of how well the model economy mimics historical data…. The
degree of confidence in the answer depends on the confidence that is
placed in the economic theory being used.” But anyone who believes that
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
34
INTRODUCTION
theories must be warranted by evidence has little reason to date to place
much confidence in real business cycle models.
NOTES
1 As will become clear below, our focus is on real business cycles narrowly construed
as perfectly competitive representative agent models driven by real shocks. A
number of recent developments have extended models with roots in Kydland
and Prescott (1982 [3]) to include monetary factors, limited heterogeneity among
agents, non-Walrasian features, and imperfect competition. These models are
ably surveyed in chapters 7–9 of Cooley (1995b). One way to view this literature
is as a constructive response to some of the difficulties with the narrow real
business cycle model that we evaluate.
2 Lucas (1978:242).
3 Friedman’s monetarist model is distinguished from Lucas’s new classical
monetary model in that Friedman imagines that people can be systematically
mistaken about the true state of real wages for relatively long periods, while
Lucas argues that people have rational expectations (i.e., they make only
unsystematic mistakes) and, therefore, correct their judgments about the real
wage quickly.
4 Lucas (in Snowden, Vane and Wynarczyk, 1994:222) accepts that his previous
characterization of the Austrians as precursors to new classical business cycle
theory was incorrect.
5 Equation (1.1) is a snapshot of the economy at a particular time. In fact, variables
in the model are growing. We could indicate this with subscripts indexing the
relevant time, but this would simply clutter the notation unnecessarily.
6 There is a large literature on endogenous growth models (see, e.g., the symposium
in the Journal of Economic Perspectives, 1994).
7 Cooley and Prescott (1995, sec. 4) discusses the issues related to establishing an
appropriate correspondence between the real business cycle model and the national
accounts to permit the calibration of the model.
8 It is actually a debated question whether microeconomic studies do in fact provide
the necessary parameters. Prescott (1986a [4]: 14) cites Lucas’s (1980:712)
argument that we have “a wealth of inexpensively available data” of this sort.
However, Hansen and Heckman (1996 [14]: 93–94) argue that in this regard
Prescott is wrong. As evidence they point to Shoven and Whalley (1992:105),
who rather candidly admit that “it is surprising how sparse (and sometimes
contradictory) the literature is on some key elasticity values. And although this
procedure might sound straightforward, it is often exceedingly difficult because
each study is different from every other.” (Cf. the debate between Summers
(1986 [5]) and Prescott (1986b [6]) about whether the parameters used in Prescott
(1986a [4]) are the appropriate ones.)
9 Details on how to solve these sorts of models are provided in Chapter 2.
10 Despite this argument, Lucas’s view of real business cycle models is rather
favorable. See, e.g., the discussion in Manuelli and Sargent (1988).
11 Although we refer to z as “the technology shock,” this terminology is not universal.
Generally, z will be a persistent process; for example, zt=␳zt+εt, with ␳>0 and εt an
independent, identically distributed random variable. Some economists identify
εt as “the technology shock.” Similarly, some economists identify zt rather than Zt
as the “Solow residual.”
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE LIMITS OF BUSINESS CYCLE RESEARCH
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
35
These reasons are elaborated in Hartley (1997).
The classic reference is Gorman (1953); the literature is summarized in Stoker
(1993).
Coleman (1990) calls this the micro-to-macro transition and provides an extensive
and illuminating discussion about what is involved in making this transition
properly.
This seems to be Friedman’s (1997:210) point when he criticizes Kydland and
Prescott’s (1996 [13]) standards of empirical evaluation for calibrated models,
saying, “There is a world of difference between mimicking and explaining, between
‘can or may’ and ‘does’.”
The official definition of a business cycle used by the National Bureau of Economic
Research in the United States emphasizes that recessions are unfavorable
movements across a breadth of economic aggregates; see Geoffrey Moore quoted
in Hall (1981).
Prescott (1986a [4]: 10) argues that the noun “business cycle” should be avoided
as it encourages people to believe incorrectly that there is an entity to be explained,
independently of economic growth, which is characterized by a deterministic
cycle. Instead, “business cycle” should be used as an adjective, as in “business
cycle phenomena,” that points to the volatility and comovements of various
economic series. Lucas (1987, sec. V) acknowledges the difference between
defining the business cycle, as common usage does, as recurrent fluctuations in
unemployment and, as he and real business cycle models typically do, as
equilibrium comovements. He recognizes that, to the extent that one is interested
in questions of unemployment, models that aim to explain the comovements
alone are silent on an important question—although he argues that this is a
limitation, not a fault.
Harvey and Jaeger (1993) present evidence that the HP filter which is used to
detrend the series analyzed by Kydland and Prescott (1990 [20]) distorts the
correlations among them, suggesting that the “facts” might be artifacts of the
statistical processing after all (see section II (v) below).
Farmer (1993) is an exception, see section II (ii) below.
Both the term “stylized facts” and the facts themselves are due to Kaldor (1961).
Also see Kaldor (1957) in which the facts themselves are discussed with the
name “stylized.”
Manuelli and Sargent (1988) criticize the real business cycle literature for backing
away from following procedures along these lines.
Kydland and Prescott (1990 [20]) and Burns and Mitchell (1946).
See Fair (1990) for a description of the model.
A strategy for the assessment of idealized models is discussed in Hoover (1994).
This claim is controversial. Romer (1986a, b; 1989) argues that postwar business
cycles are not substantially less variable than those of the nineteenth century.
Weir (1986) and Balke and Gordon (1989) challenge Romer’s revisionism. The
debate is updated and assessed in Siegler (1997), which, on the basis of better
estimates of nineteenth-century GNP, supports the traditional view that
modern business cycles are in fact smoother than those of the nineteenth
century.
Solow (1956 [27]: 314, 320) explicitly observes that idle capacity biases the measure
and that the measure hinges on the assumption of factors being paid their
marginal products, but that a similar measure could be created for monopolistic
factor markets. Looking back over thirty years later, Solow (1990:225) argues
that he never intended the Solow residual as a suitable measure of anything but
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
36
INTRODUCTION
the trend in technology: “the year-to-year behavior of the residual could be
governed by all sorts of ‘technologically’ irrelevant short-run forces. I still think
that….”
27 The HP filter is defined as follows: Let xt=x-t + x^t where x-t denotes the trend
component and x^t denotes the deviation from trend. The HP filter chooses this
decomposition to solve the following problem:
28
29
30
31
32
T is the number of observations and λ is a parameter that controls the amount of
smoothness in the series: if λ=0, then the smooth series is identical to the
original series; if λ=∞, the smoothed series is just a linear trend. Hodrick and
Prescott use a value of λ=1600 for quarterly data on the ground that this replicates
the curve a business cycle analyst might fit freehand to the data. With no better
justification than this, λ=1600 has become the standard choice for the smoothing
parameter in the real business cycle literature.
For example, Prescott (1986a [4]: 16) argues against comparing the model output
to the path of actual US output.
Christiano (1988) seems to have been the first real business cycle modeler to
employ such a test.
Additionally, independently detrending the Solow residual and the other inputs
to the production function may introduce discrepancies between actual and
model-generated data.
Hoover and Salyer (1996) provide simulation evidence that the Solow residual
does not convey useful information about technology shocks, and that the
apparent success of real business cycle models in matching historical data for
output is wholly an artifact of the use of current output in the construction of
the Solow residual.
If, for instance, we were to condition on actual output, inherited capital and the
government expenditure shock, then we could back out another measure of zS.
But, given that we have nothing independent to compare it with, the more
interesting point is that we can back out some other series, say, labor conditional
on actual output and capital, which can then be compared to its actual counterpart.
REFERENCES
Adelman, Irma and Frank L.Adelman (1959) “The Dynamic Properties of
the Klein-Goldberger Model,” Econometrica, 27(4): 596–625.
Altu, Sumru (1989) “Time-to-Build and Aggregate Fluctuations: Some New
Evidence,” International Economic Review, 30(4), November: 889–920,
reprinted here in Chapter 21.
Backus, David K. and Patrick J.Kehoe (1992) “International Evidence on the
Historical Properties of Business Cycles,” American Economic Review, 82(4):
864–88.
Balke, Nathan S. and Robert J.Gordon (1989) “The Estimation of Prewar
Gross National Product: Methodology and New Evidence,” Journal of Political
Economy, 97(1): 38–92.
Benhabib, Jess, Richard Rogerson, and Randall Wright (1991) “Homework in
Macroeconomics: Household Production and Aggregate Fluctuations,”
Journal of Political Economy, 99(6), December: 1166–87.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE LIMITS OF BUSINESS CYCLE RESEARCH
37
Bryant, John (1983) “A Simple Rational Expectations Keynes-type Model,”
Quarterly Journal of Economics, 98(3), August: 525–28.
Burns, Arthur F. and Wesley C.Mitchell (1946) Measuring Business Cycles, New
York: National Bureau of Economic Research.
Calomiris, Charles W. and Christopher Hanes (1995) “Historical
Macroeconomics and Macroeconomic History,” in Kevin D.Hoover (ed.)
Macroeconometrics: Developments, Tensions, and Prospects, Dordrecht: Kluwer:
351–416.
Canova, Fabio (1991) “Detrending and Business Cycle Facts,” unpublished
typescript, Department of Economics, European University Institute,
Florence, Italy, July 20.
Canova, Fabio, M.Finn, and A.R.Pagan (1994) “Evaluating a Real Business
Cycle Model,” in C.Hargreaves (ed.) Nonstationary Time Series Analysis and
Cointegration, Oxford: Oxford University Press: 225–55, reprinted here
in Chapter 22.
Christiano, Lawrence J. (1988) “Why Does Inventory Investment Fluctuate
So Much?” Journal of Monetary Economics, 21(2): 247–80.
Christiano, Lawrence J. and Martin Eichenbaum (1992) “Current Real Business
Cycle Theories and Aggregate Labor Market Fluctuations,” American Economic
Review, 82(3), June: 430–50, reprinted here in Chapter 10.
Cogley, Timothy and James M.Nason (1995a) “Effects of the Hodrick-Prescott
Filter on Trend and Difference Stationary Time Series: Implications for
Business Cycle Research,” Journal of Economic Dynamics and Control, 19(1–
2), January-February: 253–78, reprinted here in Chapter 33.
Cogley, Timothy and James M.Nason (1995b) “Output Dynamics in Real
Business Cycle Models,” American Economic Review, 85(3), June: 492–511,
reprinted here in Chapter 30.
Coleman, James S. (1990) Foundations of Social Theory, Cambridge, Mass.: The
Belknap Press.
Cooley, Thomas F. (1995a) “Contribution to a Conference Panel Discussion:
What Do We Know about How Monetary Policy Affects the Economy?”
paper given at the 19th Annual Economic Policy Conference, Federal
Reserve Bank of St Louis. Federal Reserve Bank of St Louis Review, 77(3):
131–37.
Cooley, Thomas F. (ed.) (1995b) Frontiers of Business Cycle Research, Princeton:
Princeton University Press.
Cooley, Thomas F. and Gary D.Hansen (1989) “The Inflation Tax in a Real
Business Cycle Model,” American Economic Review, 79(4), September: 733–
48, reprinted here in Chapter 11.
Cooley, Thomas F. and Edward C.Prescott (1995) “Economic Growth and
Business Cycles,” in Cooley (1995b): 1–38.
Cooper, Russell and Andrew John (1988) “Coordinating Coordination Failures,”
Quarterly Journal of Economics, 103(3), August: 441–64.
Debreu, Gerard (1974) “Excess Demand Functions,” Journal of Mathematical
Economics, 1(1), March: 15–23.
Eichenbaum, Martin (1991) “Real Business Cycle Theory: Wisdom or
Whimsy?” Journal of Economic Dynamics and Control, 15(4), October: 607–
26, reprinted here in Chapter 25.
Fair, Ray C. (1990) Fairmodel: User’s Guide and Intermediate Workbook, Southborough, Mass.: Macro Incorporated.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
38
INTRODUCTION
Farmer, Roger E.A. (1993) The Macroeconomics of Self-fulfilling Prophecies, Cambridge,
Mass.: MIT Press.
Friedman, Milton (1957) A Theory of the Consumption Function, Princeton: Princeton
University Press.
Friedman, Milton (1968) “The Role of Monetary Policy,” American Economic
Review, 58(1): 1–17.
Friedman, Milton (1997) “Computational Experiments,” Journal of Economic
Perspectives, 11(1), Winter: 209–10.
Frisch, Ragnar (1933) “Propagation Problems and Impulse Response Problems
in Dynamic Economics,” in Economic Essays in Honour of Gustav Cassel: October
20, 1933, London: George Allen & Unwin.
Gorman, W.M. (1953) “Community Preference Fields,” Econometrica, 21(1): 63–80.
Greenwood, Jeremy, Zvi Hercowitz, and Gregory W.Huffman (1988)
“Investment, Capacity Utilization and the Real Business Cycle,” American
Economic Review, 78(3), June: 402–17.
Gregory, Allan W. and Gregor W.Smith (1990) “Calibration as Estimation,”
Econometric Reviews, 9(1): 57–89.
Gregory, Allan W. and Gregor W.Smith (1991) “Calibration as Testing: Inference
in Simulated Macroeconomic Models,” Journal of Business and Economic
Statistics, 9(3), July: 297–303, reprinted here in Chapter 16.
Griliches, Zvi (1996) “The Discovery of the Residual: A Historical Note,” Journal
of Economic Literature, 34(3): 1324–30, reprinted here in Chapter 29.
Hall, Robert E. (1981) “Just a Peak and Trough,” National Bureau of Economic
Research (downloaded from World Wide Web site: nber.harvard.edu).
Hall, Robert E. (1986) “Market Structure and Macroeconomics Fluctuations,”
Brookings Papers on Economic Activity, no. 2:265–338.
Hall, Robert E. (1990) “Invariance Properties of Solow’s Productivity Residual,”
in Peter Diamond (ed.) Growth/Productivity/Unemployment: Essays to Celebrate
Bob Solow’s Birthday, Cambridge, Mass.: MIT Press: 71–112.
Hansen, Gary D. (1985) “Indivisible Labor and the Business Cycle,” Journal of
Monetary Economics, 16(3), November: 309–28, reprinted here in Chapter 8.
Hansen, Gary D. and Edward C.Prescott (1993) “Did Technology Shocks
Cause the 1990–1991 Recession?” American Economic Review, 83(2), May:
280–86, reprinted here in Chapter 26.
Hansen, Gary D. and Thomas J.Sargent (1988) “Straight Time and Overtime
in Equilibrium,” Journal of Monetary Economics, 21(2), March: 281–308.
Hansen, Gary D. and Randall Wright (1992) “The Labor Market in Real
Business Cycle Theory,” Federal Reserve Bank of Minneapolis Quarterly Review,
16(2), Spring: 2–12, reprinted here in Chapter 9.
Hansen, Lars Peter and James J.Heckman (1996) “The Empirical Foundations
of Calibration,” Journal of Economic Perspectives, 10(1), Winter: 87–104,
reprinted here in Chapter 14.
Hansen, Lars Peter and Thomas J.Sargent (1980) “Formulating and Estimating
Dynamic Linear Rational Expectations Models,” Journal of Economic Dynamics
and Control, 2(1), February, reprinted in Robert E.Lucas, Jr. and Thomas J.
Sargent (eds) (1981) Rational Expectations and Econometric Practice, London:
George Allen & Unwin: 91–125.
Hansen, Lars Peter and Thomas J.Sargent (1990) “Recursive Linear Models
of Dynamic Economies,” National Bureau of Economic Research, Working
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE LIMITS OF BUSINESS CYCLE RESEARCH
39
Paper Series, no. 3479.
Hartley, James E. (1994) “Technology in Macroeconomic Models,” doctoral
dissertation, University of California, Davis.
Hartley, James E. (1997) The Representative Agent in Macroeconomics, London:
Routledge.
Hartley, James E., Kevin D.Salyer, and Steven M.Sheffrin (1997) “Calibration
and Real Business Cycle Models: An Unorthodox Experiment,” Journal
of Macroeconomics, 19(1): 1–17, reprinted here in Chapter 24.
Harvey, A.C. and A.Jaeger (1993) “Detrending, Stylized Facts and the Business
Cycle,” Journal of Applied Econometrics, 8(3): 231–47, reprinted here in
Chapter 32.
Hayek, Friedrich A. von (1933) Monetary Theory and the Trade Cycle, trans. Nicholas
Kaldor and H.M.Croome, London: Jonathan Cape.
Hayek, Friedrich A. von (1935) Prices and Production, 2nd ed., London: Routledge
& Kegan Paul.
Hodrick, Robert J. and Edward C.Prescott (1997) “Postwar US Business Cycles:
An Empirical Investigation,” Journal of Money, Credit, and Banking, 29(1),
February: 1–16, reprinted here in Chapter 31.
Hoover, Kevin D. (1988) The New Classical Economics: A Sceptical Inquiry, Oxford:
Blackwell.
Hoover, Kevin D. (1994) “Six Queries About Idealization in an Empirical
Context,” Poznan Studies in the Philosophy of Science and the Humanities, 38:43–
53.
Hoover, Kevin D. (1995) “Facts and Artifacts: Calibration and the Empirical
Assessment of Real-Business-Cycle Models,” Oxford Economic Papers, 47(1),
March: 24–44, reprinted here in Chapter 15.
Hoover, Kevin D. (1997) “Comments on Smith and Zin’s ‘Real Business Cycle
Realizations, 1925–1995’,” Carnegie-Rochester Series in Public Policy, forthcoming.
Hoover, Kevin D. and Kevin D.Salyer (1996) “Technology Shocks or Colored
Noise? Why Real Business Cycle Models Cannot Explain Actual Business
Cycles,” unpublished manuscript.
Hume, David ([1739] 1888) A Treatise of Human Nature, Oxford: Clarendon
Press.
Ingrao, Bruna and Giorgio Israel (1990) The Invisible Hand: Economic Equilibrium
in the History of Science, trans. Ian McGilvray, Cambridge, Mass.: MIT Press.
Jorgenson, Dale W. and Griliches, Zvi (1967) “The Explanation of Productivity
Change,” Review of Economic Studies, 34(3), July: 249–83.
Journal of Economic Perspectives (1994) Symposium on “New Growth Theory,”
8(1): 3–72.
Kaldor, Nicholas (1957) “A Model of Economic Growth,” Economic Journal,
67(4), December: 591–624.
Kaldor, Nicholas (1961) “Capital Accumulation and Economic Growth,” in
Friedrich A.Lutz and Douglas C.Hague (eds) The Theory of Capital: Proceedings
of a Conference Held by the International Economics Association, London: Macmillan.
King, Robert G. (1995) “Quantitative Theory and Econometrics,” Federal Reserve
Bank of Richmond Economic Quarterly, 81(3), Summer: 53–105.
King, Robert G. and Charles I.Plosser (1989) “Real Business Cycles and the
Test of the Adelmans,” Journal of Monetary Economics, 33(2), April: 405–38,
reprinted here in Chapter 23.
King, Robert G., Charles I.Plosser and Sergio T.Rebelo (1988) “Production,
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
40
INTRODUCTION
Growth, and Business Cycles I: The Basic Neoclassical Model,” Journal of
Monetary Economics, 21(2), March: 195–232, reprinted here in Chapter 7.
Kirman, Alan P. (1989) “The Intrinsic Limits of Modern Economic Theory:
The Emperor Has No Clothes,” The Economic Journal, 99 (Conference):
126–39.
Kirman, Alan P. (1992) “Whom or What Does the Representative Individual
Represent?” Journal of Economic Perspectives, 6(2), Spring: 117–36.
Kydland, Finn E. and Edward C.Prescott (1982) “Time to Build and Aggregate
Fluctuations,” Econometrica, 50(6), November: 1345–69, reprinted here in
Chapter 3.
Kydland, Finn E. and Edward C.Prescott (1988) “The Workweek of Capital
and Its Cyclical Implications,” Journal of Monetary Economics, 21(2): 343–
60.
Kydland, Finn E. and Edward C.Prescott (1990) “Business Cycles: Real Facts
and a Monetary Myth,” Federal Reserve Bank of Minneapolis Quarterly Review,
14(2), Spring: 3–18, reprinted here in Chapter 20.
Kydland, Finn E. and Edward C.Prescott (1991) “The Econometrics of the
General Equilibrium Approach to Business Cycles,” Scandinavian Journal
of Economics, 93(2): 161–78, reprinted here in Chapter 12.
Kydland, Finn E. and Edward C.Prescott (1996) “The Computational
Experiment: An Econometric Tool,” Journal of Economic Perspectives, 10(1),
Winter: 69–86, reprinted here in Chapter 13.
Kydland, Finn E. and Edward C.Prescott (1997) “A Response [to Milton
Friedman],” Journal of Economic Perspectives, 11(1), Winter: 210–11.
Lucas, Robert E., Jr. (1972) “Expectations and the Neutrality of Money,” Journal
of Economic Theory , 4(2), April: 103–24, reprinted in Lucas (1981): 66–89.
Lucas, Robert E., Jr. (1973) “Some Output-Inflation Tradeoffs,” American Economic
Review, 63(3), June: 326–34, reprinted in Lucas (1981): 131–45.
Lucas, Robert E., Jr. (1975) “An Equilibrium Model of the Business Cycle,”
Journal of Political Economy, 83(6): 1113–44, reprinted in Lucas (1981):
179–214.
Lucas, Robert E., Jr. (1976) “Econometric Policy Evaluation: A Critique,” in
Karl Brunner and Allan H.Meltzer (eds) The Phillips Curve and Labor Markets,
vol. 1 of Carnegie-Rochester Conference Series on Public Policy,
Amsterdam: North Holland, 19–46, reprinted in Lucas (1981): 104–30.
Lucas, Robert E., Jr. (1977) “Understanding Business Cycles,” in Karl Brunner
and Allan H.Meltzer (eds) Stabilization of the Domestic and International Economy,
Carnegie-Rochester Conference Series in Public Policy, Amsterdam: North
Holland, 7–29, reprinted in Lucas (1981): 215–39.
Lucas, Robert E., Jr. (1978) “Unemployment Policy,” American Economic Review,
68(2), May: 353–57, reprinted in Lucas (1981): 240–47.
Lucas, Robert E., Jr. (1980) “Methods and Problems in Business Cycle Theory,”
Journal of Money, Credit and Banking, 12(4), pt. 2, November: 696–713,
reprinted in Lucas (1981): 271–96.
Lucas, Robert E., Jr. (1981) Studies in Business Cycle Theory, Cambridge, Mass.:
MIT Press.
Lucas, Robert E., Jr. (1987) Models of Business Cycles, Oxford: Blackwell.
Lucas, Robert E., Jr., and Thomas J.Sargent (1979) “After Keynesian
Macroeconomics,” Federal Reserve Bank of Minneapolis Quarterly Review, 3(2),
reprinted in Robert E.Lucas, Jr. and Thomas J.Sargent (eds) (1981) Rational
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE LIMITS OF BUSINESS CYCLE RESEARCH
41
Expectations and Econometric Practice, Minneapolis: The University of Minnesota
Press, 295–319.
Mankiw, N.Gregory (1989) “Real Business Cycles: A New Keynesian
Perspective,” Journal of Economic Perspectives, 3(3), Summer: 79–90, reprinted
here in Chapter 28 .
Mantel, R. (1974) “On the Characterization of Aggregate Excess Demand,”
Journal of Economic Theory, 7(3): 348–53.
Mantel, R. (1976) “Homothetic Preferences and Community Excess Demand
Functions,” Journal of Economic Theory, 12(2): 197–201.
Manuelli, Rodolfo and Thomas J.Sargent (1988) “Models of Business Cycles: A
Review Essay,” Journal of Monetary Economics, 22:523–42.
Mas-Collel, A. (1977) “On the Equilibrium Price Set of an Exchange Economy,”
Journal of Mathematical Economics, 4(2): 117–26.
Nelson, Charles R. and H.Kang (1981) “Spurious Periodicity in Inappropriately
Detrended Time Series,” Econometrica, 49(3), May: 741–51.
Popper, Karl (1959) The Logic of Scientific Discovery, London: Hutchinson.
Popper, Karl (1972) Conjectures and Refutations: The Growth of Scientific Knowledge,
4th ed., London: Routledge.
Prescott, Edward C. (1986a) “Theory Ahead of Business Cycle Measurement,”
in Federal Reserve Bank of Minneapolis Quarterly Review, 10(4), Fall: 9–22,
reprinted here in Chapter 4.
Prescott, Edward C. (1986b) “Response to a Skeptic,” Federal Reserve Bank of
Minneapolis Quarterly Review, 10(4), Fall: 28–33, reprinted here in Chapter
6.
Romer, Christina (1986a) “Is the Stabilization of the Postwar Economy a
Figment of the Data?” American Economic Review, 76(3), June: 314–34.
Romer, Christina (1986b) “New Estimates of Prewar Gross National Product
and Unemployment,” Journal of Economic History, 46:341–52.
Romer, Christina (1989) “The Prewar Business Cycle Reconsidered: New
Estimates of Gross National Product, 1869–1908,” Journal of Political
Economy, 97(1), February: 1–37.
Samuelson, Paul A. (1948) Economics: An Introductory Analysis, New York: McGrawHill.
Sargent, Thomas J. (1979) Macroeconomic Theory, New York: Academic Press.
Sato, R. (1966) “On the Adjustment Time in Neo-Classical Growth Models,”
Review of Economic Studies, 33, July: 263–68.
Shafer, Wayne and Hugo Sonnenschein. (1982) “Market Demand and Excess
Demand Functions,” in K.J.Arrow and M.D.Intriligator (eds) Handbook of
Mathematical Economics, vol. 2, Amsterdam: North-Holland: 671–93.
Shoven, John B. and John Whalley (1992) Applying General Equilibrium, New
York: Cambridge University Press.
Siegler, Mark V. (1997) “Real Output and Business Cycle Volatility, 1869–
1993: US Experience in International Perspective,” doctoral dissertation,
University of California, Davis.
Slutsky, Eugen E. ([1927] 1937) “The Summation of Random Causes as the
Source of Cyclic Processes”, Econometrica, 5:105–46. Originally published
in Russian in 1927.
Smith, Gregor and Stanley Zin (1997) “Real Business Cycle Realizations, 1925–
1995,” Carnegie-Rochester Series in Public Policy, forthcoming.
Snowdon, Brian, Howard Vane, and Peter Wynarczyk (eds) (1994) A Modern
Guide to Macroeconomics, Aldershot: Edward Elgar.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
42
INTRODUCTION
Solow, Robert M. (1956) “A Contribution to the Theory of Economic Growth,”
Quarterly Journal of Economics, 70(1), February: 65–94.
Solow, Robert M. (1957) “Technical Change and the Aggregate Production
Function,” Review of Economics and Statistics, 39(3), August: 312–20, reprinted
here in Chapter 27.
Solow, Robert M. (1970) Growth Theory: An Exposition, Oxford: Blackwell.
Solow, Robert M. (1990) “Reactions to Conference Papers,” in Peter Diamond
(ed.) Growth/Productivity/Unemployment: Essays to Celebrate Bob Solow’s Birthday,
Cambridge, Mass.: MIT Press: 221–29.
Sonnenschein, Hugo (1973) “Do Walras’ Law and Continuity Characterize
the Class of Community Excess Demand Functions?” Journal of Economic
Theory, 6(4): 345–54.
Sonnenschein, Hugo (1974) “Market Excess Demand Functions,” Econometrica,
40(3), May: 549–63.
Stoker, Thomas M. (1993) “Empirical Approaches to the Problem of
Aggregation over Individuals,” Journal of Economic Literature, 21(4),
December: 1827–74.
Summers, Lawrence H. (1986) “Some Skeptical Observations on Real Business
Cycle Theory,” Federal Reserve Bank of Minneapolis Quarterly Review, 10(4),
Fall: 23–27, reprinted here in Chapter 5.
Watson, Mark W. (1993) “Measures of Fit for Calibrated Models,” Journal of
Political Economy, 101(6), December: 1011–41, reprinted here in Chapter
17.
Weir, David (1986) “The Reliability of Historical Macroeconomic Data for
Comparing Cyclical Stability,” Journal of Economic History, 46(2), June:
353–65.
Yule, G.Udny (1926) “Why Do We Sometimes Get Nonsense-Correlations
between Time-Series?” Journal of the Royal Statistical Society, 89(1): 1–65.
Reprinted in David F.Hendry and Mary S.Morgan (eds) The Foundations of
Econometric Analysis. Cambridge: Cambridge University Press, 1995, ch. 9.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
Chapter 2
A user’s guide to solving real business
cycle models
The typical real business cycle model is based upon an economy populated
by identical infinitely lived households and firms, so that economic choices
are reflected in the decisions made by a single representative agent. It is
assumed that both output and factor markets are characterized by perfect
competition. Households sell capital, kt, to firms at the rental rate of capital,
and sell labor, ht, at the real wage rate. Each period, firms choose capital
and labor subject to a production function to maximize profits. Output is
produced according to a constant-returns-to-scale production function
that is subject to random technology shocks. Specifically yt=ztf(kt, ht), where
yt is output and zt is the technology shock. (The price of output is normalized
to one.) Households’ decisions are more complicated: given their initial
capital stock, agents determine how much labor to supply and how much
consumption and investment to purchase. These choices are made in
order to maximize the expected value of lifetime utility. Households must
forecast the future path of wages and the rental rate of capital. It is assumed
that these forecasts are made rationally. A rational expectations equilibrium
consists of sequences for consumption, capital, labor, output, wages, and
the rental rate of capital such that factor and output markets clear.
While it is fairly straightforward to show that a competitive equilibrium
exists, it is difficult to solve for the equilibrium sequences directly. Instead,
an indirect approach is taken in which the Pareto optimum for this
economy is determined (this will be unique given the assumption of
representative agents). As shown by Debreu (1954), the Pareto optimum
as characterized by the optimal sequences for consumption, labor, and
capital in this environment will be identical to that in a competitive
equilibrium. Furthermore, factor prices are determined by the marginal
products of capital and labor evaluated at the equilibrium quantities. (For
a detailed exposition of the connection between the competitive equilibrium
and Pareto optimum in a real business cycle model, see Prescott, 1986
[4].) We now provide an example of solving such a model.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
44
INTRODUCTION
I DERIVING THE EQUILIBRIUM CONDITIONS
The first step in solving for the competitive equilibrium is to determine
the Pareto optimum. To do this, the real business cycle model is recast as
the following social planner’s problem:
(2.1)
where E1[·] denotes expectations conditional on information at t=1, 0<β
<1 is agents’ discount factor, ct denotes consumption, (1-ht) is leisure (agents’
endowment of time is normalized to one), it is investment, and 0<δ<1 is
the depreciation rate of capital. The exogenous technology shock is
assumed to follow the autoregressive process given in the last equation;
the autocorrelation parameter is 0ⱕρⱕ1 and the innovation to technology
is assumed to have a mean of one and standard deviation σε. The first two
constraints in equation (2.1) are the economy-wide resource constraint,
and the second is the law of motion for the capital stock.
Dynamic programming problem
This infinite horizon problem can be solved by exploiting its recursive
structure. That is, the nature of the social planner’s problem is the same
every period: given the beginning-of-period capital stock and the current
technology shock, choose consumption, labor, and investment. Note that
utility is assumed to be time-separable: that is, the choices of consumption
and labor at time t do not affect the marginal utilities of consumption and
leisure in any other time period. Because of this recursive structure, it is
(2.2)
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
GUIDE TO SOLVING REAL BUSINESS CYCLES
45
useful to cast the maximization problem as the following dynamic
programming problem (for a discussion of dynamic programming, see
Sargent (1987)):
(Note that investment has been eliminated by using the law of motion for
the capital stock.) A solution to this problem must satisfy the following
necessary conditions and resource constraint:
Where the notation Ui,t; i=l, 2 denotes the derivative of the utility function
with respect to the ith argument evaluated at the quantities (ct, 1-ht); fi,t;
i=1,2 has an analogous interpretation. N1 represents the intratemporal
efficiency condition (the labor-leisure tradeoff). It implies that the marginal
rate of substitution between labor and consumption must equal the marginal
product of labor. The second condition, N2, represents the intertemporal
efficiency condition. The left-hand side represents the marginal cost in
terms of utility of investing in more capital, while the right-hand side
represents the expected marginal utility gain; at an optimum, these costs
and benefits must be equal.
To simplify the analysis (again, see Prescott (1986 [4]) for a justification),
assume the following functional forms:
(The assumption that utility is linear in leisure is based on Hansen’s
(1985 [8]) model.) Then the three equilibrium conditions become:
(2.3)
A steady-state equilibrium for this economy is one in which the
technology shock is assumed to be constant, so that there is no uncertainty:
that is, zt=1 for all t, and the values of capital, labor, and consumption are
for all t. Imposing these steady-state
constant,
conditions in equation (2.3), the steady-state values are found by solving
the following steady-state equilibrium conditions:
In the above expressions, denotes the steady-state level of output.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
46
INTRODUCTION
Calibration
The next step in solving the model is to choose parameter values for the
model. This is done through calibration: the set of parameters (δ, ß, A, α)
are chosen so that the steady-state behavior of the model matches the
long-run characteristics of the data. The features of the data which do not
exhibit cyclical characteristics are:
1
2
3
4
(1-α)=labor’s average share of output.
ß-1-1=average risk-free real interest rate.
Given (α, ß) choose d so that the output-capital ratio (from (SS2) is
consistent with observation.
The parameter A determines the time spent in work activity. To see
this, multiply both sides of (SS1) by and rearrange the expression to
But the steady-state resource constraint,
yield:
so that the output-consump(SS3), implies that
tion ratio is implied by the parameter values chosen in the previous
three steps. Hence, the choice of A directly determines
Typical parameter values based on postwar US data (see Hansen and
Wright (1992 [9])) are: α=0.36 implying labor’s share is 64 percent, ß=0.99
implying an annual riskless interest rate of 0.04 percent, δ=0.025 implying
the capital-output ratio (where output is measured on a quarterly basis) of
roughly 10, and A=3 which implies that roughly 30 percent of time is
spent in work activity. (These values will be used in Section IV below.)
II LINEARIZATION
The solution to the social planner’s problem is characterized by a set of
policy functions for capital, consumption, and labor; moreover, the solution
exists and is unique (see Prescott (1986 [4])). There is, however, no
analytical solution. To make the model operational, therefore, an
approximate numerical solution is found. One of the simplest methods is
to take a linear approximation (i.e., a first-order Taylor series expansion)
of the three equilibrium conditions and the law of motion of the technology
Provided the stochastic
shock around the steady-state values
behavior of the model does not push the economy too far from the steadystate behavior, the linear approximation will be a good one. (The discussion
below follows closely that of Farmer (1993).) This technique is demonstrated
below.1
Intratemporal efficiency condition
The optimal labor-leisure choice is represented by condition N1:
Linearizing around the steady-state values :
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
GUIDE TO SOLVING REAL BUSINESS CYCLES
47
Note that in the last expression, all variables have been expressed as
percentage deviations from the steady state (the first two terms modify the
in steady
respective derivatives while the last term uses the fact that
state). Consumption can be expressed as a percentage deviation from steady
condition dividing
state by using the steady-state
both sides of the equation by this expression and denoting percentage
deviations from steady state as equation (2.4) can be written as:
(2.5)
Intertemporal efficiency condition
This efficiency condition is given by N2:
Again, linearizing around the steady state, and expressing all variables as
percentage deviations from steady state, yields:
Multiplying each side of the equation by and using the steady-state
condition (SS2) that
(2.6)
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
48
INTRODUCTION
Resource constraint
Following the same procedure as before, linearizing the resource constraint
around the steady state yields:
(2.7)
Technology shock process
The critical difference between the steady-state model and the real
business cycle model is the assumption that technology shocks are
random—the shocks follow the autoregressive process described in
equation (2.1). Linearizing the autoregressive process for the technology
shock results in:
(2.8)
Taking expectations of both sides:
(2.9)
III SOLUTION METHOD
The equations that define a rational expectations equilibrium (equations
(2.5), (2.6), (2.7), and (2.9)) can be written as a vector expectational
difference equation. Let
where bold print denotes a vector, then the linear system of equations can
be written as:
(2.10)
The matrices A and B are:
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
GUIDE TO SOLVING REAL BUSINESS CYCLES
49
Premultiplying both sides of equation (2.10) by A-1 yields:
(2.11)
The matrix A-1B can be decomposed as (see Hamilton (1994) for details):
(2.12)
where Q is a matrix whose columns are the eigenvectors of A-1B and ⌳ is
a diagonal matrix whose diagonal elements are the eigenvalues of A-1B.
Using this decomposition and premultiplying both sides of the resulting
expression in equation (2.11) by Q-1 yields:
(2.13)
Note that the elements of the defined (4×1) column vector dt are
constructed from a linear combination of the elements in the rows of the
(4× 4) matrix Q-1 and the elements of the (4×1) column vector ut. Since ⌳
is a diagonal matrix, equation (2.13) implies four independent equations:
(2.14)
Since the equations in equation (2.14) must hold every period, it is possible
to recursively substitute the expressions forward for T periods to yield:
(2.15)
The λi are four distinct eigenvalues associated with the four equilibrium
conditions (equations (2.5)–(2.8)). Since one of these conditions is the
law of motion for the exogenous technology shock (equation (2.8)), one
of the eigenvalues will be ρ-1. Also, the first rows of the matrices A and B
are determined by the intratemporal efficiency condition; since this is not
a dynamic relationship, one of the eigenvalues will be zero. The remaining
two eigenvalues will bracket the value of unity as is typical for a saddle
path equilibrium implied by the underlying stochastic growth framework.
As implied by equation (2.15), the stable, rational expectations solution to
the expectational difference equation is associated with the eigenvalue
with a value less than one. That is, if λi>1, then iterating forward implies
di,t→∞ which is not a permissible equilibrium. Furthermore, for equation
(2.15) to hold for all T (again taking the limit of the right-hand side), in
the stable case when, λ<1, it must be the true that di,t=0; this restriction
provides the desired solution. That is, di,t=0 imposes the linear restriction
which is consistent with a rational expectations solution.
on
(Recall that di,t represents a linear combination between the elements of a
particular row of Q-1 and the elements of the vector ut.)
IV A PARAMETRIC EXAMPLE
In this section, a parameterized version of the RBC model described
above is solved. The following parameter values are used: (ß=0.99,
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
50
INTRODUCTION
α=0.36, δ=0.025, A=3). These imply the following steady-state values:
Note that these values imply
that agents spend roughly 30 percent of their time in work activities and
the capital-output ratio is approximately 10 (output is measured on a
quarterly basis); both of these values are broadly consistent with US
experience (see McGrattan, 1994).
The remaining parameter values determine the behavior of the
technology shock. These are estimated by constructing the Solow residual2
and then detrending that series linearly. Specifically, the Solow residual
is defined as Zt=1n yt-α 1n kt-(1-α) 1n ht. The Zt series can then be regressed
on a linear time trend (which is consistent with the assumption of constant
technological progress) and the residual is identified as the technology
shock zt. Using this procedure on quarterly data over the period 60.1-94.4
resulted in an estimate of the serial correlation of zt (the parameter ?) to be
0.95. The variance of the shock to technology (i.e., the variance of et in
equation (2.8)) was estimated to be 0.007. Note that the variance of the
technology shock is not relevant in solving the linearized version of the
model—however, when the solution of the model is used to generate
artificial time series in the simulation of the economy, this parameter value
must be stipulated.
These values generated the following entries into the A and B matrices:
Following the steps described in the previous section (premultiplying by
A-1) yields the following:
Next, decomposing A-1 B into Q⌳
⌳ Q-1 and then premultiplying by Q-1
yields
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
GUIDE TO SOLVING REAL BUSINESS CYCLES
51
The entries in the matrix ⌳ (i.e., the eigenvalues of A-1B) determine the
solution. Note that the second diagonal entry is (accounting for rounding
error) ρ-1. The fourth row of ⌳ is associated with the intratemporal efficiency
condition. These values are proportional to those given in the first row of the
A matrix; consequently, dividing all entries by (-2.62) returns the original
intratemporal efficiency condition. The remaining two entries in the ⌳ matrix
are those related to the saddle path properties of the steady-state solution.
Since a stable rational expectations solution is associated with an eigenvalue
less than unity, the third row of the Q-1 matrix provides the linear restriction
we are seeking. That is, the rational expectations solution is:
(2.16)
The law of motion for the capital stock (the parameter values are given in
the third row of the A matrix) and the intratemporal efficiency condition
provides two more equilibrium conditions:
(2.17)
(2.18)
A random number generator can next be used to produce a sequence of
technology shocks. The above equilibrium equations can then be used to
produce time series for capital, consumption, labor, and output.
V ANALYZING OUTPUT FROM THE ARTIFICIAL
ECONOMY
The solution to the model is characterized by equations (2.16)–(2.18).
Given initial values for capital, and next generating a path for the exogenous
these equations will produce time series for
technology shock
Two other series that most macroeconomists are interested in,
namely, output and investment, can be generated by linearizing the
production function and the resource constraint, respectively.
Specifically, for output, linearizing the assumed Cobb-Douglas
and using the calibrated value
production function (i.e.,
that α= 0.36) yields the following equation:
(2.19)
Finally, a linear approximation of the condition that, in equilibrium, output
must equal the sum of consumption and investment, can be expressed in
the form of a percentage deviation from the steady state as:
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
52
INTRODUCTION
(2.20)
Using the steady-state values employed in the numerical solution, the
investment equation becomes:
(2.21)
Hence, equilibrium in this economy is described by the following set of
equations:
To generate the time series implied by the model, it is necessary first to
generate a series for the innovations to the technology shock, i.e., These
Figure 2.1 Output, Consumption, and Investment in RBC Model
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
GUIDE TO SOLVING REAL BUSINESS CYCLES
53
Table 2.1 Descriptive Statistics for US and RBC model
Source: Statistics for US data are taken from Kydland and Prescott (1990 [20]), tables
I and II, pp. 10–11.
are assumed to have a mean of zero and a variance that is consistent with
the observed variance for the innovations, which, as mentioned above, is
and using a random number
roughly 0.007. Then, initializing
generator in order to generate the innovations, a path for the technology
shocks is created. Next, assuming that all remaining values are initially at
their steady state (which implies that all initial values are set to zero), the
system of equations above can be solved to produce the time path for the
endogenous variables.
We generate artificial time paths for consumption, output, and
investment (3000 observations were created and only the last 120 were
examined), and these are shown in Figure 2.1. It is clear from Figure 2.1,
as is also true in the actual data, that the volatility of investment is greater
than that of output, which is greater than that of consumption. To see this
more precisely, the standard deviation of consumption, labor, and
investment relative to output is reported in Table 2.1, along with the
correlations of these series with output.
NOTES
1
2
Recall that the general form for the Taylor series expansion of a function around
a point x* is:
where N! denotes factorial.
The use of the Solow residual as a measure of technology shocks is discussed in
Hoover and Salyer (1996).
REFERENCES
Debreu, Gerard (1954) “Valuation Equilibrium and Pareto Optimum,”
Proceedings of the National Academy of Science, 40:588–92.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
54
INTRODUCTION
Farmer, Roger E.A. (1993) The Macroeconomics of Self-fulfilling Prophecies, Cambridge,
Mass.: MIT Press.
Hamilton, James D. (1994) Time Series Analysis, Princeton: Princeton University
Press.
Hansen, Gary D. (1985) “Indivisible Labor and the Business Cycle,” Journal of
Monetary Economics, 16(3), November: 309–28, reprinted here in Chapter 8.
Hansen, Gary D. and Randall Wright (1992) “The Labor Market in Real
Business Cycle Theory,” Federal Reserve Bank of Minneapolis Quarterly Review,
16(2), Spring: 2–12, reprinted here in Chapter 9.
Hoover, Kevin D. and Kevin D.Salyer (1996) “Technology Shocks or Colored
Noise? Why Real Business Cycle Models Cannot Explain Actual Business
Cycles,” unpublished manuscript.
Kydland, Finn E. and Edward C.Prescott (1990) “Business Cycles: Real Facts
and a Monetary Myth,” Federal Reserve Bank of Minneapolis Quarterly Review,
14(2), Spring: 3–18, reprinted here in Chapter 20.
McGratten, Ellen R. (1994) “A Progress Report on Business Cycle Models,”
Federal Reserve Bank of Minneapolis Quarterly Review, 18(4), Fall: 2–16.
Prescott, Edward C. (1986) “Theory Ahead of Business Cycle Measurement,”
Federal Reserve Bank of Minneapolis Quarterly Review, 10(4), Fall: 9–
22, reprinted here in Chapter 4.
Sargent, Thomas J. (1987) Dynamic Macroeconomic Theory, Cambridge, Mass.:
Harvard University Press.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
Part II
The foundations of real business
cycle modeling
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
CHAPTER 3
57
ECONOMETRICA
VOLUME 50
NOVEMBER, 1982
NUMBER 6
TIME TO BUILD AND AGGREGATE FLUCTUATIONS
BY FINN E.KYDLAND AND EDWARD C.PRESCOTT1
The equilibrium growth model is modified and used to explain the cyclical variances
of a set of economic time series, the covariances between real output and the other
series, and the autocovariance of output. The model is fitted to quarterly data for the
post-war U.S. economy. Crucial features of the model are the assumption that more
than one time period is required for the construction of new productive capital, and
the non-time-separable utility function that admits greater intertemporal substitution
of leisure. The fit is surprisingly good in light of the model’s simplicity and the small
number of free parameters.
1. INTRODUCTION
THAT WINE IS NOT MADE in a day has long been recognized by economists
(e.g., Böhm–Bawerk [6]). But, neither are ships nor factories built in a day. A thesis
of this essay is that the assumption of multiple-period construction is crucial for
explaining aggregate fluctuations. A general equilibrium model is developed and
fitted to U.S. quarterly data for the post-war period. The co-movements of the
fluctuations for the fitted model are quantitatively consistent with the corresponding
co-movements for U.S. data. In addition, the serial correlations of cyclical output for
the model match well with those observed.
Our approach integrates growth and business cycle theory. Like standard growth
theory, a representative infinitely-lived household is assumed. As fluctuations in
employment are central to the business cycle, the stand-in consumer values not only
consumption but also leisure. One very important modification to the standard
growth model is that multiple periods are required to build new capital goods and
only finished capital goods are part of the productive capital stock. Each stage of
production requires a period and utilizes resources. Half-finished ships and factories
are not part of the productive capital stock. Section 2 contains a short critique of the
commonly used investment technologies, and presents evidence that single-period
production, even with adjustment costs, is inadequate. The preference-technologyinformation structure of the model is presented in Section 3. A crucial feature of
preferences is the non-time-separable utility function that admits greater intertemporal
substitution of leisure. The exogenous stochastic components in the model are shocks
to technology and imperfect indicators of productivity. The two technology shocks
differ in their persistence.
The steady state for the model is determined in Section 4, and quadratic
approximations are made which result in an “indirect” quadratic utility function that
1
The research was supported by the National Science Foundation. We are grateful to Scan
Becketti, Fischer Black, Robert S.Chirinko, Mark Gersovitz, Christopher A.Sims, and John B.
Taylor for helpful comments, to Sumru Altug for research assistance, and to the participants in the
seminars at the several universities at which earlier drafts were presented.
1345
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
58
FOUNDATIONS
1346
F.E.KYDLAND AND E.C.PRESCOTT
values leisure, the capital goods, and the negative of investments. Most of the
relatively small number of parameters are estimated using steady state considerations.
Findings in other applied areas of economics are also used to calibrate the model.
For example, the assumed number of periods required to build new productive
capital is of the magnitude reported by business, and findings in labor economics are
used to restrict the utility function. The small set of free parameters imposes
considerable discipline upon the inquiry. The estimated model and the comparison
of its predictions with the empirical regularities of interest are in Section 5. The final
section contains concluding comments.
2. A CRITIQUE OF CONVENTIONAL AGGREGATE INVESTMENT
TECHNOLOGIES
There are two basic technologies that have been adopted in empirical studies of
aggregate investment behavior. The first assumes a constant-returns-to-scale
neoclassical production function F with labor L and capital K as the inputs. Total
output F(K, L) constrains the sum of investment and consumption, or C+I≤F(K,
L), where C, I, K, L≥0. The rate of change of capital,
is investment less
depreciation, and depreciation is proportional with factor δ to the capital stock,
that is,
This is the technology underlying the work of Jorgenson
[19] on investment behavior.
An implication of this technology is that the relative price of the investment and
consumption goods will be a constant independent of the relative outputs of the two
goods.2 It also implies that the shadow price of existing capital will be the same as
the price of the investment good.3 There is a sizable empirical literature that has
found a strong association between the level of investment and a shadow price of
capital obtained from stock market data (see [26]). This finding is inconsistent with
this assumed technology as is the fact that this shadow price varies considerably
over the business cycle.
The alternative technology, which is consistent with these findings, is the single
capital good adjustment cost technology.4 Much of that literature is based upon the
problem facing the firm and the aggregation problem receives little attention. This has
led some to distinguish between internal and external adjustment costs. For aggregate
investment theory this is not an issue (see [29]) though for other questions it will be.
Labor resources are needed to install capital whether the acquiring or supplying firm
installs the equipment. With competitive equilibrium it is the aggregate production
2
This, of course, assumes neither C nor I is zero. Sargent [32], within a growth context with
shocks to both preferences and technology, has at a theoretical level analyzed the equilibrium with
corners. Only when investment was zero did the price of the investment good relative to that of the
consumption good become different from one and then it was less than one. This was not an
empirical study and Sargent states that there currently are no computationally practical econometric
methods for conducting an empirical investigation within that theoretical framework.
3
The shadow price of capital has been emphasized by Brunner and Meltzer [7] and Tobin [36]
in their aggregate models.
4
See [1, 17] for recent empirical studies based on this technology.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
TIME TO BUILD
AGGREGATE FLUCTUATIONS
59
1347
possibility set that matters. That is, if the Yj are the production possibility sets of the
firms associated with a given industrial organization and for some other industrial
organization, the same aggregate supply behavior results if
The adjustment cost model, rather than assuming a linear product transformation
curve between the investment and consumption goods, imposes curvature. This can
be represented by the following technology:
where G like F is increasing, concave, and homogeneous of degree one. Letting the
price of the consumption good be one, the price of the investment good qt, the
rental price of capital rt, and the wage rate wt, the firm’s problem is to maximize real
profits, Ct+qtIt-wtLt-rtKt, subject to the production constraint. As constant returns to
scale are assumed, the distribution of capital does not matter, and one can proceed
as if there were a single price-taking firm. Assuming an interior solution, given that
this technology displays constant returns to scale and that the technology is separable
between inputs and outputs, it follows that It=F(Kt, Lt)h(qt)=Zth(qt), where Zt is defined
to be aggregate output. The function h is increasing, so high investment-output
ratios are associated with a high price of the investment good relative to the
consumption good. Figure 1 depicts the investment-consumption product
transformation curve and Figure 2 the function h(q). For any I/Z, the negative of the
slope of the transformation curve in Figure 1 is the height of the curve in Figure 2.
This establishes that a higher q will be associated with higher investment for this
technology. This restriction of the theory is consistent with the empirical findings
previously cited.
There are other predictions of this theory, however, which are questionable. If
we think of the q-investment curve h depicted in Figure 2 as a supply curve, the
short- and the long-run supply elasticities will be equal. Typically, economists argue
that there are specialized resources which cannot be instantaneously and costlessly
transferred between industries and that even though short-run elasticities may be
low, in the long run supply elasticities are high. As there are no specialized resources
for the adjustment cost technology, such considerations are absent and there are no
penalties resulting from rapid adjustment in the relative outputs of the consumption
and investment good.
FIGURE 1.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
FIGURE 2.
60
FOUNDATIONS
1348
F.E.KYDLAND AND E.C.PRESCOTT
To test whether the theory is a reasonable approximation, we examined crosssection state data. The correlations between the ratios of commercial construction
to either state personal income or state employment and price per square foot5
are both -0.35. With perfectly elastic supply and uncorrelated supply and demand
errors, this correlation cannot be positive. To explain this large negative
correlation, one needs a combination of high variability in the cross-sectional
supply relative to cross-sectional demand plus a positive slope for the supply
curve. Our view is that, given mobility of resources, it seems more plausible
that the demand is the more variable. Admitting potential data problems, this
cross-sectional result casts some doubt upon the adequacy of the single capital
good adjustment cost model.
At the aggregate level, an implication of the single capital good adjustment
cost model is that when the investment-output ratio is regressed on current and
lagged q, only current q should matter.6 The findings in [26] are counter to this
prediction.
In summary, our view is that neither the neoclassical nor the adjustment cost
technologies are adequate. The neoclassical structure is inconsistent with the
positive association between the shadow price of capital and investment activity.
The adjustment cost technology is consistent with this observation, but inconsistent
with cross-sectional data and the association of investment with the lagged as
well as the current capital shadow prices. In addition, the implication that longand short-run supply elasticities are equal is one which we think a technology
should not have.
Most destructive of all to the adjustment-cost technology, however, is the finding
that the time required to complete investment projects is not short relative to the
business cycle. Mayer [27], on the basis of a survey, found that the average time
(weighted by the size of the project) between the decision to undertake an investment
project and the completion of it was twenty-one months. Similarly, Hall [13] found
the average lag between the design of a project and when it becomes productive to
be about two years. It is a thesis of this essay that periods this long or even half that
long have important effects upon the serial correlation properties of the cyclical
components of investment and total output as well as on certain co-movements of
aggregate variables.
The technological requirement that there are multiple stages of production is not
the delivery lag problem considered by Jorgenson [19]. He theorized at the firm
level and imposed no consistency of behavior requirement for suppliers and
demanders of the investment good. His was not a market equilibrium analysis and
there was no theory accounting for the delivery lag. Developing such a market
theory with information asymmetries, queues, rationing, and the like is a challenging
problem confronting students of industrial organization.
5
The data on commercial construction and price per square foot were for 1978 and were
obtained from F.W.Dodge Division of McGraw-Hill.
6
This observation is due to Fumio Hayashi.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
TIME TO BUILD
AGGREGATE FLUCTUATIONS
61
1349
Our technology assumes that a single period is required for each stage of
construction or that the time required to build new capital is a constant. This is not
to argue that there are not alternative technologies with different construction periods,
patterns of resource use, and total costs. We have found no evidence that the capital
goods are built significantly more rapidly when total investment activity is higher or
lower. Lengthening delivery lags (see [9]) in periods of high activity may be a matter
of longer queues and actual construction times may be shorter. Premiums paid for
earlier delivery could very well be for a more advanced position in the queue than
for a more rapidly constructed factory. These are, of course, empirical questions,
and important cyclical variation in the construction period would necessitate an
alternative technology.
Our time-to-build technology is consistent with short-run fluctuations in the shadow
price of capital because in the short run capital is supplied inelastically. It also implies
that the long-run supply is infinitely elastic, so on average the relative price of the
investment good is independent of the investment-output ratio.
3. THE MODEL
Technology
The technology assumes time is required to build new productive capital. Let Sjt be
the number of projects j stages or j periods from completion for j =1, …, J-1, where
J periods are required to build new productive capacity. New investment projects
initiated in period t are SJt. The recursive representation of the laws of motion of
these capital stocks is
(3.1)
(3.2)
Here, kt is the capital stock at the beginning of period t, and δ is the depreciation rate.
The element SJt is a decision variable for period t.
The final capital good is the inventory stock yt inherited from the previous period.7
Thus, in this economy, there are J+1 types of capital: inventories yt, productive
capital kt, and the capital stocks j stages from completion for j=1, …, J-1. These
variables summarize the effects of past decisions upon current and future production
possibilities.
Let j for j=1, …, J be the fraction of the resources allocated to the
investment project in the jth stage from the last. Total non-inventory investment
in period t is
Total investment, i t, is this amount plus inventory
7
All stocks are beginning-of-the-period stocks.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
62
FOUNDATIONS
1350
F.E.KYDLAND AND E.C.PRESCOTT
investment yt+1-yt, and consequently
(3.3)
Total output, that is, the sum of consumption ct and investment, is constrained as
follows:
(3.4)
where nt is labor input, t a shock to technology, and f is a constant-returns-to-scale
production function to be parameterized subsequently.
Treating inventories as a factor of production warrants some discussion. With
larger inventories, stores can economize on labor resources allocated to restocking.
Firms, by making larger production runs, reduce equipment down time associated
with shifting from producing one type of good to another. Besides considerations
such as these, analytic considerations necessitated this approach. If inventories were
not a factor of production, it would be impossible to locally approximate the economy
using a quadratic objective and linear constraints. Without such an approximation
no practical computational method currently exists for computing the equilibrium
process of the model.
The production function is assumed to have the form
(3.5)
where 0<<1, 0<σ<1, and 0<<∞. This form was selected because, among other
things, it results in a share for labor in the steady state. The elasticity of substitution
between capital and inventory is 1/(1+ν). This elasticity is probably less than one
which is why ν is required to be positive.
Preferences
The preference function, whose expected value the representative household
maximizes, has the form
where 0<β<1 is the discount factor, lt
leisure, L the lag operator, and
Normalizing so that one is the
endowment of time, we let nt=1-lt be the time allocated to market activity. The
polynomial lag operator is restricted so that the αi sum to one, and αi=(1-η)i-1α1 for
i≥1, where 0<η≤1. With these restrictions,
By defining the variable
recursive representation:
the distributed lag has the following
(3.6)
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
TIME TO BUILD
AGGREGATE FLUCTUATIONS
63
1351
The variable at summarizes the effects of all past leisure choices on current and
future preferences. If ns=nt for all s≤t, then at=nt/η, and the distributed lag is
simply 1-nt.
The parameters α0 and determine the degree to which leisure is intertemporally
substitutable. We require 0<η≤1 and 0<α0≤1. The nearer a0 is to one, the less is
the intertemporal substitution of leisure. For α0 equal to one, time-separable utility
results. With η equal to one, at equals nt-1. This is the structure employed in [33].
As η approaches zero, past leisure choices have greater effect upon current utility
flows.
Non-time-separable utility functions are implicit in the empirical study of aggregate
labor supply in [25]. Grossman [12] and Lucas [24] discuss why a non-time-separable
utility function is needed to explain the business cycle fluctuations in employment
and consumption. A micro justification for our hypothesized structure based on a
Beckerian household production function is as follows.8 Time allocated to non-market
activities, that is lt, is used in household production. If there is a stock of household
projects with varying output per unit of time, the rational household would allocate
lt to those projects with the greatest returns per time unit. If the household has
allocated a larger amount of time to non-market activities in the recent past, then
only projects with smaller yields should remain. Thus, if at is lower, the marginal
utility value of lt should be smaller.
Cross-sectional evidence of households’ willingness to redistribute labor supply
over time is the lumpiness of that supply. There are vacations and movements of
household members into and out of the labor force for extended periods which are
not in response to large movements in the real wage. Another observation suggesting
high intertemporal substitutability of leisure is the large seasonal variation in hours
of market employment. Finally, the failure of Abowd and Ashenfelter [2] to find a
significant wage premium for jobs with more variable employment and earnings
patterns is further evidence. In summary, household production theory and crosssectional evidence support a non-time-separable utility function that admits greater
intertemporal substitution of leisure—something which is needed to explain aggregate
movements in employment in an equilibrium model.
The utility function in our model is assumed to have the form
where γ<1 and γ0. If the term in the square brackets is interpreted as a
composite commodity, then this is the constant-relative-risk-aversion utility function
with the relative degree of risk aversion being 1-γ. We thought this composite
commodity should be homogeneous of degree one as is the case when there is a
single good. The relative size of the two exponents inside the brackets is
8
We thank Nasser Saïdi for suggesting this argument.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
64
FOUNDATIONS
1352
F.E.KYDLAND AND E.C.PRESCOTT
motivated by the fact that households’ allocation of time to nonmarket activities is
about twice as large as the allocation to market activities.
Information Structure
We assume that the technology parameter is subject to a stochastic process with
components of differing persistence. The productivity parameter is not observed
but the stand-in consumer does observe an indicator or noisy measure of this
parameter at the beginning of the period. This might be due to errors in reporting
data or just the fact that there are errors in the best or consensus forecast of what
productivity will be for the period. On the basis of the indicator and knowledge of
the economy-wide state variables, decisions of how many new investment projects
to initiate and of how much of the time endowment to allocate to the production of
marketed goods are made. Subsequent to observing aggregate output, the consumption
level is chosen with inventory investment being aggregate output less fixed investment
and consumption.
Specifically, the technology shock, λt, is the sum of a permanent component, λ1t,
and a transitory component,9 λ2t:
(3.7)
In the spirit of the Friedman-Muth permanent-income model, the permanent
component is highly persistent so
(3.8)
where ρ is less than but near one and ζ1t is a permanent shock.10 The transitory
component equals the transitory shock so
(3.9)
The indicator of productivity, πt, is the sum of actual productivity λt and a third
shock ζ3t:
(3.10)
The shock vectors ζt=(ζ1t, ζ2t, ζ3t) are independent multivariate normal with mean
vector zero and diagonal covariance matrix.
The period-t labor supply decision nt and new investment project decision sJt are
made contingent upon the past history of productivity shocks, the λk for k<t, the
indicator of productivity πt, the stocks of capital inherited from the past, and variable
9
The importance of permanent and transitory shocks in studying macro fluctuations is emphasized
in [8].
10
The value used for ρ in this study was 0.95. The reason we restricted ρ to be strictly less than
one was technical. The theorem we employ to guarantee the existence of competitive equilibrium
requires stationarity of the shock.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
TIME TO BUILD
AGGREGATE FLUCTUATIONS
65
1353
at. These decisions cannot be contingent upon λt for it is not observed or deducible
at the time of these decisions. The consumption-inventory investment decision,
however, is contingent upon λt for aggregate output is observed prior to this decision
and λt can be deduced from aggregate output and knowledge of inputs.
The state space is an appropriate formalism for representing this recursive
information structure. Because of the two-stage decision process, it is not a direct
application of Kalman filtering. Like that approach the separation of estimation and
control is exploited. The general structure assumes an unobservable state vector,
say xt, that follows a vector autoregressive process with independent multivariate
normal innovations:
(3.11)
Observed prior to selecting the first set of decisions is
(3.12)
The element B1 is a matrix and the e1t are independent over time. Observed prior to
the second set of decisions and subsequent to the first set is
(3.13)
Equations (3.11)–(3.13) define the general information structure.
To map our information structure into the general formulation, let
B1=[1 1], B2=[1 1],
V1=[var(3)], and V2=[0]. With these definitions, the information structure (3.7)–
(3.10) viewed as deviations from the mean and the representation (3.11)– (3.13)
are equivalent.
Let m0t be the expected value and 0 the covariance of the distribution of xt
conditional upon the pk=(p1k, p2k) for k<t. Using the conditional probability laws for
the multivariate normal distribution (see [28, p. 208]) and letting m1t and 1 be the
mean and covariance of xt conditional upon p1t as well, we obtain
(3.14)
(3.15)
Similarly, the mean vector m2t and covariance matrix 2 conditional upon p2t as well
are
(3.16)
(3.17)
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
66
FOUNDATIONS
1354
F.E.KYDLAND AND E.C.PRESCOTT
Finally, from (3.11),
(3.18)
(3.19)
The covariances 0, 1, and 2 are defined recursively by (3.15), (3.17), and (3.19).
The matrix V0 being of full rank along with the stability of A are sufficient to insure
that the method of successive approximations converges exponentially fast to a
unique solution.
The covariance elements 0, 1, and 2 do not change over time and are therefore
not part of the information set. The m0t, m1t, and m2t do change but are sufficient
relative to the relevant histories for forecasting future values of both the unobserved
state and the observable pt, >t, and for estimating the current unobserved state.
Equilibrium
To determine the equilibrium process for this model, we exploit the well-known
result that, in the absence of externalities, competitive equilibria are Pareto optima.
With homogeneous individuals, the relevant Pareto optimum is the one which
maximizes the welfare of the stand-in consumer subject to the technology constraints
and the information structure. Thus, the problem is to
subject to constraints (3.1)–(3.4), (3.6), and (3.11)–(3.13), given k0, s10, …, sJ-1,0, a0,
and that x0~N(m0, 0). The decision variables at time t are nt, sJt, ct, and yt+1. Further,
nt and sJt cannot be contingent upon p2t for it is observed subsequent to these
decisions.
This is a standard discounted dynamic programming problem. There are optimal
time-invariant or stationary rules of the form
nt=n(kt, s1t, s2t, …, sJ-1,t, yt, at, m1t),
sJt=s(kt, s1t, s2t, …, SJ-1,t, yt, at, m1t),
ct=c(kt, s1t, s2t, …, sJt, yt, at, nt, m2t),
yt+1=y(kt, s1t, s2t, …, sJt, yt, at, nt, m2t).
It is important to note that the second pair of decisions are contingent upon m2t
rather than m1t and that they are contingent also upon the first set of decisions sJt
and nt.
The existence of such decision rules and the connection with the competitive
allocation is established in [31]. But, approximations are necessary before equilibrium
decision rules can be computed. Our approach is to determine the steady
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
TIME TO BUILD
AGGREGATE FLUCTUATIONS
67
1355
state for the model with no shocks to technology. Next, quadratic approximations
are made in the neighborhood of the steady state. Equilibrium decision rules for the
resulting approximate economy are then computed. These rules are linear, so in
equilibrium the approximate economy is generated by a system of stochastic difference
equations for which covariances are easily determined.
4. STEADY STATE, APPROXIMATION, AND COMPUTATION OF
EQUILIBRIUM
Variables without subscript denote steady state values. The steady state interest rate
is r=(1- )/ , and the steady state price of (non-inventory) capital
The latter is obtained by observing that 1 units of
consumption must be foregone in the current period, 2 units the period before,
etc., in order to obtain one additional unit of capital for use next period.
Two steady state conditions are obtained by equating marginal products to rental
rates, namely fy=r and fk=q(r+). These imply fk/fy=q(r+)/r. For production function
(3.5), this reduces to
(4.1)
Differentiating the production function with respect to capital, substituting for y
from (4.1), and equating to the steady-state rental price, one obtains
where
Solving for k as a function of n yields
(4.2)
Steady-state output as a function of n is
steady state, net investment is zero, so
In the
(4.3)
The steady-state values of c, k, and y are all proportional to
We also note that
the capital-output ratio is b3/b4, and that consumption’s share to total steady-state
output is 1-(b3/b4).
Turning now to the consumer’s problem and letting µ be the Lagrange multiplier
for the budget constraint and wt the real wage, first-order conditions are
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
68
FOUNDATIONS
1356
F.E.KYDLAND AND E.C.PRESCOTT
In the steady state, ct=c, lt=l, and wt=w for all t. Making these substitutions and using
the fact that the i sum to one, these expressions simplify to
Eliminating µ from these equations yields
and l=1-n, this in turn implies
Since
(4.4)
Returning to the production side, the marginal product of labor equals the real
wage:
(4.5)
Using (4.3) and (4.5), we can solve (4.4) for n:
That n does not depend upon average matches well with the American experience
over the last thirty years. During this period, output per man-hour has increased by
a few hundred per cent, yet man-hours per person in the 16–65 age group has
changed but a few per cent.
Approximation About the Steady State
If the utility function u were quadratic and the production function f linear, there
would be no need for approximations. In equilibrium, consumption must be equal
to output minus investment. We exploit this fact to eliminate the nonlinearity in the
constraint set by substituting f(, k, n, y)-i for c in the utility function to obtain u(f(,
k, n, y)-i, n, a). The next step is to approximate this function by a quadratic in the
neighborhood of the model’s steady state. As investment i is linear in the decision
and state variables, it can be eliminated subsequent to the approximation and still
preserve a quadratic objective.
Consider the general problem of approximating function u(x) near x-. The
approximate quadratic function is
where x, b ∈ Rn and Q is an n×n symmetric matrix. We want an approximation
that is good not only at x- but also at other x in the range experienced during the
sample period. Let zi be a vector, all of whose components are zero except for
Our approach is to select the elements bi and qii so that the approximation
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
TIME TO BUILD
AGGREGATE FLUCTUATIONS
69
1357
error is zero at the
and
where the
selected correspond to the
approximate average deviations of the xi from their steady state values
The
used for , k, y, n, i, and a were 3, 1, 2, 3, 8, and 0.5 per cent,
values of
respectively.11
The approximation errors being zero at the
and
requires that
The elements qij, i ≠j, are selected to minimize the sum of the squared approximation
errors at
and
The approximation
error at the first point is
Summing over the square of this error and the three others, differentiating with
respect to qij, setting the resulting expression equal to zero and solving for qij, we
obtain
for i ≠j.
Computation of Equilibrium
The equilibrium process for the approximate economy maximizes the welfare of
the representative household subject to the technological and informational constraints
as there are no externalities. This simplifies the determination of the equilibrium
process by reducing it to solving a linear-quadratic maximization problem. For such
mathematical structure there is a separation of estimation and control. Consequently,
the first step in determining the equilibrium decision rules for the approximate
economy is to solve the following deterministic problem:
11
We experimented a little and found that the results were essentially the same when the
second order Taylor series approximation was used rather than this function. Larry
Christiano [10] has found that the quadratic approximation method that we employed
yields approximate solutions that are very accurate, even with large variability, for a structure
that, like ours, is of the constant elasticity variety.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
70
FOUNDATIONS
1358
F.E.KYDLAND AND E.C.PRESCOTT
subject to
(4.6)
(4.7)
(4.8)
(4.9)
(4.10)
(4.11)
At this stage, the fact that there is an additive stochastic term in the equation
determining xt+1 is ignored as is the fact that xt is not observed for our economy.
Constraints (4.6)–(4.9) are the laws of motion for the state variables. The free
decision variables are nt, sJt, and yt+1. It was convenient to use inventories taken into
the subsequent period, yt+1, as a period t decision variable rather than it because the
decisions on inventory carry-over and consumption are made subsequent to the
labor supply and new project decisions nt and SJt.
For notational simplicity we let the set of state variables other than the unobserved
xt be z t=(kt, yt, at , s1t , . . . , sJ-1,t ) and the set of decision variables d t =(n t , s Jt , y t+1 ). The
unobserved state variables xt=(x1t , x2t) are the permanent and transitory shocks to
technology. Finally, v(x, z) is the value of the deterministic problem if the initial state
is (x, z). It differs from the value function for the stochastic problem by a constant.
Using constraints (4.10) and (4.11) to substitute for it and λt in the utility function,
an indirect utility function U(x, z, d) is obtained. The value function, v(x, z), was
computed by the method of successive approximations or value iteration. If vj(x, z)
is the jth approximation, then
subject to constraints (4.6)–(4.9). The initial approximation, v0(x, z), is that function
which is identically zero.
The function U is quadratic and the constraints are linear. Then, if vj is quadratic,
υ j+1 must be quadratic. As υ 0 is trivially quadratic, all the υ j are quadratic and
therefore easily computable. We found that the sequence of quadratic functions
converged reasonably quickly.12
12
The limit of the sequence of value functions existed in every case and, as a function of
z, was bounded from above, given x. This, along with the stability of the matrix A, is
sufficient to ensure that this limit is the optimal value function and that the associated policy
function is the optimal one (see [30]).
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
TIME TO BUILD
AGGREGATE FLUCTUATIONS
71
1359
The next step is to determine the optimal inventory carry-over decision rule. It is
the linear function yt+1=y(xt, zt, nt, sJt) which solves
(4.12)
subject to (4.6)–(4.9) and both nt and SJt given. Finally, the solution to the program
where v2 is thevalue of maximization of (4.12), is determined. The linear functions
SJt=s(xt, zt) and nt=n(xt, zt) which solve the above program are the optimal decision
rules for new projects and labor supply.
Because of the separation of estimation and control in our model, these decision
rules can be used to determine the motion of the stochastic economy. In each period
t, a conditional expectation, m0t, is formed on the basis of observations in previous
periods. An indicator of the technology shock is observed, which is the sum of a
permanent and a transitory component as well as an indicator shock. The conditional
expectation, m1t, of the unobserved xt is computed according to equation (3.14), and
SJt and nt are determined from
(4.13)
(4.14)
where xt has been replaced by m 1t. Then the technology shock, t, is observed, which
changes the conditional expectation of xt. From (3.16), this expectation is m2t, and the
inventory carry-over is determined from
(4.15)
To summarize, the equilibrium process governing the evolution of our economy is
given by (3.1)–(3.3), (3.6), (3.11)–(3.14), (3.16), (3.18), and (4.13)–(4.15).
5. TEST OF THE THEORY
The test of the theory is whether there is a set of parameters for which the model’s
co-movements for both the smoothed series and the deviations from the smoothed
series are quantitatively consistent with the observed behavior of the corresponding
series for the U.S. post-war economy. An added requirement is that the parameters
selected not be inconsistent with relevant micro observations, including reported
construction periods for new plants and cross-sectional observations on consumption
and labor supply. The closeness of our specification of preferences and technology
to those used in many applied studies facilitates such comparisons.
The model has been rigged to yield the observations that smoothed output,
investment, consumption, labor productivity, and capital stocks all vary roughly
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
72
FOUNDATIONS
1360
F.E.KYDLAND AND E.C.PRESCOTT
proportionately while there is little change in employment (all variables are in perhousehold terms) when the technology parameter grows smoothly over time.
These are just the steady state properties of the growth model with which we
began.
Quantitatively explaining the co-movements of the deviations is the test of the
underlying theory. For want of better terminology, the deviations will be referred to
as the cyclical components even though, with our integrated approach, there is no
separation between factors determining a secular path and factors determining
deviations from that path. The statistics to be explained are the covariations of
the cyclical components. They are of interest because their behavior is stable and is
so different from the corresponding covariations of the smoothed series. This is
probably why many have sought separate explanations of the secular and cyclical
movements.
One cyclical observation is that, in percentage terms, investment varies three
times as much as output does and consumption only half as much. In sharp contrast
to the secular observations, variations in cyclical output are principally the result of
variations in hours of employment per household and not in capital stocks or labor
productivity.
The latter observation is a difficult one to explain. Why does the consumption of
market produced goods and the consumption of leisure move in opposite directions
in the absence of any apparent large movement in the real wage over the so-called
cycle? For our model, the real wage is proportional to labor’s productivity, so the
crucial test is whether most of the variation in cyclical output arises from variations
in employment rather than from variations in labor’s productivity.
We chose not to test our model versus the less restrictive vector autoregressive
model.13 This most likely would have resulted in the model being rejected, given the
measurement problems and the abstract nature of the model. Our approach is to
focus on certain statistics for which the noise introduced by approximations and
measurement errors is likely to be small relative to the statistic. Failure of the theory
to mimic the behavior of the post-war U.S. economy with respect to these stable
statistics with high signal-noise ratios would be grounds for its rejection.
Model Calibration
There are two advantages of formulating the model as we did and then
constructing an approximate model for which the equilibrium decision rules
are linear. First, the specifications of preferences and technology are close to
those used in many applied studies. This facilitates checks of reasonableness of
many parameter values. Second, our approach facilitates the selection of parameter
values for which the model steady-state values are near average values for the
American economy during the period being explained. These two considerations
13
Sims [34] has estimated unrestricted aggregate vector autoregressive models.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
TIME TO BUILD
AGGREGATE FLUCTUATIONS
73
1361
reduce dramatically the number of free parameters that will be varied when searching
for a set that results in cyclical covariances near those observed. In explaining the
covariances of the cyclical components, there are only seven free parameters, with
the range of two of them being severely constrained a priori.
Capital for our model reflects all tangible capital, including stocks of plant and
equipment, consumer durables and housing. Consumption does not include the
purchase of durables but does include the services from the stock of consumer
durables. Different types of capital have different construction periods and patterns
of resource requirements. The findings summarized in Section 2 suggest an average
construction period of nearly two years for plants. Consumer durables, however,
have much shorter average construction periods. Having but one type of capital, we
assume, as a compromise, that four quarters are required, with one-fourth of the
value put in place each quarter. Thus J=4 and 1=2 =3=4=0.25.
Approximately ten per cent of national income account GNP is the capital
consumption allowance and another ten per cent excise tax. To GNP should be
added the depreciation of consumer durables which has the effect of increasing the
share of output going to owners of capital. In 1976, compensation to employees plus
proprietary income was approximately 64 per cent of GNP plus consumer durables
depreciation less indirect business tax, while owners of capital received about 36 per
cent. As labor share is , we set =0.64.
Different types of capital depreciate more rapidly than others, with durables
depreciating more rapidly than plant and housing, and land not depreciating at all.
As a compromise, we set the depreciation rate equal to 10 per cent per year. We
assume a subjective time discount rate of four per cent and abstract from growth.
This implies a steady-state capital to annual output ratio of 2.4. Of total output 64
per cent is wages, 24 per cent depreciation, and 12 per cent return on capital which
includes consumer durables.
The remaining parameters of technology are average , which we normalize to
one by measuring output in the appropriate units, and parameters and , which
determine the shares of and substitution between inventories and capital. Inventories
are about one-fourth of annual GNP so we require and to be such that k/y=10.
A priori reasoning indicates the substitution opportunities between capital and
inventory are small, suggesting that should be considerably larger than zero. We
restricted it to be no less than two, but it is otherwise a free parameter in our search
for a model to explain the cyclical covariances and autocovariances of aggregate
variables. Given and the value of b1=y/k, is implied. From (4.1) it is
For purposes of explaining the covariances of the
percentage deviation from steady state values, is the only free parameter associated
with technology.
The steady state real interest rate r is related to the subjective time discount
rate,
and the risk aversion parameter, , by the equation r=+ (1-)
(c/c), where c/c is the growth rate of per capita consumption. We have assumed
is four per cent per year (one per cent per quarter). As the growth rate
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
74
FOUNDATIONS
1362
F.E.KYDLAND AND E.C.PRESCOTT
of per capita consumption has been about two per cent and the real return on
physical capital six to eight per cent, the risk aversion parameter, , is constrained to
be between minus one and zero.14
The parameters 0 and which affect intertemporal substitutability of leisure
will be treated as free parameters for we could find no estimate for them in the
labor economics literature. As stated previously, the steady-state labor supply is
independent of the productivity parameter . The remaining parameters are those
specifying the process on t and the variance of the indicator. These three parameters
are var(1), var(2), and var(3). Only two of these are free parameters, however.
We restricted the sum of the three variances to be such that the estimate of the
variance of cyclical output for the model equalled that of cyclical output for the U.S.
economy during the sample period.
In summary, the parameters that are estimated from the variance-covariance
properties of the model are these variances plus the parameter determining
substitutability of inventories and capital, the parameters 0 and determining
intertemporal substitutability of leisure, and the risk aversion parameter . For each
set of parameter values, means and standard deviations were computed for several
statistics which summarize the serial correlation and covariance properties of the
model. These numbers are compared with those of the actual U.S. data for the
period 1950:1 to 1979:2 as reported in Hodrick and Prescott [18]. A set of parameter
values is sought which fits the actual data well. Having only six degrees of freedom
to explain the observed covariances imposes considerable discipline upon the analysis.
The statistics reported in [18] are not the only way to quantitatively capture the
co-movements of the deviations.15 This approach is simple, involves a minimum of
judgment, and is robust to slowly changing demographic factors which affect growth,
but are not the concern of this theory.16 In addition, these statistics are robust to
most measurement errors, in contrast to, say, the correlations between the first
differences of two series. It is important to compute the same statistics for the U.S.
economy as for the model, that is, to use the same function of the data. This is what
we do.
A key part of our procedure is the computation of dynamic competitive
equilibrium for each combination of parameter values. Because the conditional
forecasting can be separated from control in this model, the dynamic equilibrium
decision rules need only be computed for each new combination of the parameters
14
15
Estimates in [16] indicate is near zero.
With the Hodrick-Prescott method, the smooth path {st} for each series {yt} minimized
The deviations for series {yt} are {y t-s t}. The number of observations, T, was 118. The
solution to the above program is a linear transformation of the data. Thus, the standard
deviations and correlations reported are well-defined statistics.
16
See, for example, [11].
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
TIME TO BUILD
AGGREGATE FLUCTUATIONS
75
1363
TABLE I
a
MODEL PARAMETERS
a
For parameters with a time dimension, the unit of time is a quarter of a year.
, 0, , and . Similarly, the conditional expectations of the permanent and transitory
shocks which enter the decision rules depend only on the variances of the three
shocks and not upon the parameters of preferences and technology.
For each set of parameter values the following statistics are computed: the
autocorrelation of cyclical output for up to six periods, standard deviations of the
cyclical variables of interest, and their correlations with cyclical output. In [18] the
variables (except interest rates) are measured in logs while we use the levels rather
than the logs. This is of consequence only in the measurement of amplitudes, so in
order to make our results comparable to theirs, our standard deviations (except for
interest rates) are divided by the steady states of the respective variables. One can
then interpret the cyclical components essentially as percentage deviations as in
[18].
The parameter values that yielded what we considered to be the best fit are
reported in Table I. They were determined from a grid search over the free parameters.
In the case of , we tried the values 2, 3, 4, and 5. The parameters 0 and were
just constrained to be between zero and one. Only the values -1, -0.5, and -0.1 were
considered for the risk aversion parameter . The last value is close to the limiting
case of =0 which would correspond to the logarithmic utility function.
Results
All reported statistics refer to the cyclical components for both the model and the
U.S. economy. Estimated autocorrelations of real output for our model along with
sample values for the U.S. economy in the post-war period are reported in Table II.
The fit is very good, particularly in light of the model’s simplicity.
Table III contains means of standard deviations and correlations with output for
the model’s variables. Table IV contains sample values of statistics for the post-war
U.S. economy as reported in [18].
The variables in our model do not correspond perfectly to those available for the
U.S. economy so care must be taken in making comparisons. A second problem is
that there may be measurement errors that seriously bias the estimated correlations
and standard deviations. A final problem is that the estimates for the U.S. economy
are subject to sampling error. As a guide to the magnitude of this variability,
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
76
FOUNDATIONS
1364
F.E.KYDLAND AND E.C.PRESCOTT
TABLE II
a
AUTOCORRELATIONS OF OUTPUT
a
The length of the sample period both for the model and for the U.S. economy is 118 quarters.
we report the standard deviations of sample distributions for the model’s statistics
which, like the estimates for the U.S. economy, use only 118 observations. These
are the numbers in the parentheses in Tables II and III.
The model is consistent with the large (percentage) variability in investment and
low variability in consumption and their high correlations with real output. The
model’s negative correlation between the capital stock and output is consistent with
the data though its magnitude is somewhat smaller.
Inventories for our model correspond to finished and nearly finished goods
while the inventories in Table IV refer to goods in process as well. We added half
TABLE III
MODEL’S STANDARD DEVIATIONS AND CORRELATIONS WITH REAL OUTPUTa
a
b
The length of the sample period both for the model and for the U.S. economy is 118 quarters.
Measured in per cent.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
TIME TO BUILD
AGGREGATE FLUCTUATIONS
77
1365
TABLE IV
SAMPLE STANDARD DEVIATIONS AND CORRELATIONS WITH REAL
OUTPUT U.S. ECONOMY 1950:1–1979:2
the value of uncompleted capital goods to the model’s inventory variable to obtain
what we call inventories plus. This corresponds more closely to the U.S. inventory
stock variable, with its standard deviation and correlation with real output being
consistent with the U.S. data.
In Table III we include results for the implicit real interest rate given by the
expression rt=(u/ct)/( E(u/ct+l))-1. The expectation is conditional on the
information known when the allocation between consumption and inventory carryover is made.
The model displays more variability in hours than in productivity, but not by as
much as the data show. In light of the difficulties in measuring output and, in particular,
employment, we do not think this discrepancy is large. For example, all members of
the household may not be equally productive, say due to differing stocks of human
capital. If there is a greater representation in the work force of the less productive,
for example less experienced youth, when output is high, hours would be
overestimated. The effects of such errors would be to bias the variability of
employment upwards. It also would bias the correlation between productivity and
output downwards, which would result in the model being nearly consistent with the
data. Measurement errors in employment that are independent of the cycle would
have a similar effect on the correlation between output and productivity.
Another possible explanation is the oversimplicity of the model. The shocks to
technology, given our production function, are pure productivity shocks. Some
shocks to technology alter the transformation between the consumption and
investment goods. For example, investment tax credits, accelerated depreciation,
and the like, have such effects, and so do some technological changes. Further,
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
78
FOUNDATIONS
1366
F.E.KYDLAND AND E.C.PRESCOTT
some technological change may be embodied in new capital, and only after the
capital becomes productive is there the increment to measured productivity. Such
shocks induce variation in investment and employment without the variability in
productivity. This is a question that warrants further research.
We also examined lead and lag relationships and serial correlation properties of
aggregate series other than output. We found that, both for the post-war U.S.
economy and the model, consumption and non-inventory investment move
contemporaneously with output and have serial correlation properties similar to
output. Inventory and capital stocks for the model lag output, which also matches
well with the data. Some of the inventory stock’s cross-serial correlations with output
deviate significantly, however, from those for the U.S. economy. The one variable
whose lead-lag relationship does not match with the data is productivity. For the
U.S. economy it is a leading indicator, while there is no lead or lag in the model. This
was not unexpected in view of our discussion above with regard to productivity.
Thus, even though the overall fit of the model is very good, it is not surprising,
given the level of abstraction, that there are elements of the fine structure of dynamics
that it does not capture.
The Smoothed Series
The smoothed output series for the U.S. post-war data deviated significantly from
the linear time trend. During the 118-quarter sample period this difference had two
peaks and two troughs. The times between such local extremes were 30, 31, and 32
quarters, and the corresponding differences in values at adjacent extremes were
5.00, 7.25, and 5.90 per cent, respectively.
These observations match well with the predictions of the model. The mean of
the model’s sampling distribution for the number of peaks and troughs in a 118quarter period is 4.0—which is precisely the number observed. The mean of the
number of quarters between extremes is 26.1 with standard deviation 9.7, and the
mean of the vertical difference in the values at adjacent extremes is 5.0 with standard
deviation 2.9. Thus, the smoothed output series for the U.S. economy is also consistent
with the model.
Sensitivity of Results to Parameter Selection
With a couple of exceptions, the results were surprisingly insensitive to the values of
the parameters. The fact that the covariations of the aggregate variables in the
model are quite similar for broad ranges of many of the parameters suggests that,
even though the parameters may differ across economies, the nature of business
cycles can still be quite similar.
We did find that most of the variation in technology had to come from
its permanent component in order for the serial correlation properties of the model
to be consistent with U.S. post-war data. We also found that the variance of the
indicator shock could not be very large relative to the variance of the permanent
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
TIME TO BUILD
AGGREGATE FLUCTUATIONS
79
1367
technology shock. This would have resulted in cyclical employment varying less
than cyclical productivity which is inconsistent with the data.
Of particular importance for the model is the dependence of current utility on
past leisure choices which admits greater intertemporal substitution of leisure. The
purpose of this specification is not to contribute to the persistence of output
changes. If anything, it does just the opposite. This element of the model is crucial in
making it consistent with the observation that cyclical employment fluctuates
substantially more than productivity does. For the parameter values in Table I, the
standard deviation of hours worked is 18 per cent greater than the deviation of
productivity. The special case of 0=1 corresponds to a standard time-separable
utility function. For this case, with the parameters otherwise the same as in Table I,
the standard deviation of hours is 24 per cent less than the deviation of
productivity.
Importance of Time to Build
Of particular interest is the sensitivity of our results to the specification of investment
technology. The prominent alternative to our time-to-build technology is the
adjustment-cost structure. If only one period is required for the construction of new
productive capital, we can write the law of motion for the single capital good as
kt+1=(1-)kt+st, where st is the amount of investment in productive capital in period
t. We can then introduce cost of adjustment into the model by modifying the resource
constraint (3.4) as follows:
where the parameter is nonnegative. The model in Section 3 implied that the price
of investment goods, it, relative to consumption goods, ct, must be one. This price
will now of course generally not equal one, but our cost-of-adjustment formulation
insures that it is one when net investment is zero.
The magnitude of the adjustment cost can probably best be judged in terms of
the effect it has on this relative price of investment goods which differs from one by
the amount 2(st-kt). If, for example, the parameter is 0.5, and the economy is
near its steady state, a one per cent increase in the relative price of the investment
good would be associated with a four per cent increase in gross investment which is
approximately one per cent of GNP.
Even when the adjustment cost is of this small magnitude, the covariance properties
of the model are grossly inconsistent with the U.S. data for the post-war period. In
particular, most of the fluctuation of output in this model is caused by productivity
changes rather than changes in work hours. The standard deviation of hours is 0.60,
while the standard deviation of productivity is 1.29. This is just the opposite of what
the U.S. data show.
Further evidence of the failure of the cost-of-adjustment model is that, relative to
the numbers reported in Table III for our model, the standard deviation is
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
80
FOUNDATIONS
1368
F.E.KYDLAND AND E.C.PRESCOTT
nearly doubled for consumption and reduced by a factor of two for investment
expenditures, making the amplitudes of these two output components much too
close as compared with the data. In addition, the standard deviation of capital stock
was reduced by more than one half. The results were even worse for larger values
of .
The extreme case of =0 corresponds to the special case of J=1 in our model.
Thus, neither time to build nor cost of adjustment would be an element of the
model. The biggest changes in the results for this version as compared with Table III
are that the correlation between capital stock and output becomes positive and of
sizable magnitude (0.43 if the parameters are otherwise the same as in Table I), and
that the correlation between inventory stock and output becomes negative (-0.50
for our parameter values). Both of these correlations are inconsistent with the
observations. Also, the persistence of movements in investment expenditures as
measured by the autocorrelations was substantially reduced.
For our model with multiple periods required to build new capital, the results are
not overly sensitive to the number of periods assumed. With a three or five-quarter
construction period instead of four, the fit is also good.
6. CONCLUDING COMMENTS
A competitive equilibrium model was developed and used to explain the
autocovariances of real output and the covariances of cyclical output with other
aggregate economic time series for the post-war U.S. economy. The preferencetechnology environment used was the simplest one that explained the quantitative comovements and the serial correlation properties of output. These results indicate a
surprisingly good fit in light of the model’s simplicity.
A crucial element of the model that contributed to persistence of output
movements was the time-to-build requirement.17 We experimented with adjustment
costs, the standard method for introducing persistence (e.g., [4, 33]), and found that
they were not a substitute for the time-to-build assumption in explaining the data.18
One problem was that, even with small adjustment costs, employment and investment
fluctuations were too small and consumption fluctuations too large to match with
the observations.
There are several refinements which should improve the performance of the
model. In particular, we conjecture that introducing as a decision variable the hours
17
Capital plays an important role in creating persistence in the analysis of Lucas [23] as
well as in those of Blinder and Fischer [5] and Long and Plosser [22]. In [23] gradual
diffusion of information also plays a crucial role. This is not the case in our model, however,
as agents learn the value of the shock at the end of the period. Townsend [37] analyzes a
model in which decision makers forecast the forecasts of others, which gives rise to
confounding of laws of motion with forecasting problems, and results in persistence in
capital stock and output movements.
18
An alternative way of obtaining persistence is the use of long-term staggered nominal
wage contracts as in [35].
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
TIME TO BUILD
AGGREGATE FLUCTUATIONS
81
1369
per week that productive capital is employed, with agents having preferences defined
on hours worked per week, should help. Introducing more than a single type of
productive capital, with different types requiring different periods for construction
and having different patterns of resource requirement, is feasible. It would then be
possible to distinguish between plant, equipment, housing, and consumer durables
investments. This would also have the advantage of permitting the introduction of
features of our tax system which affect transformation opportunities facing the
economic agents (see, e.g., [14]). Another possible refinement is in the estimation
procedure. But, in spite of the considerable advances recently made by Hansen and
Sargent [15], further advances are needed before formal econometric methods can
be fruitfully applied to testing this theory of aggregate fluctuations.
Models such as the one considered in this paper could be used to predict the
consequence of a particular policy rule upon the operating characteristics of the
economy.19 As we estimate the preference-technology structure, our structural
parameters will be invariant to the policy rule selected even though the behavioral
equations are not. There are computational problems, however, associated with
determining the equilibrium behavioral equations of the economy when feedback
policy rules, that is, rules that depend on the aggregate state of the economy, are
used. The competitive equilibrium, then, will not maximize the welfare of the standin consumer, so a particular maximization problem cannot be solved to find the
equilibrium behavior of the economy. Instead, methods such as those developed in
[20] to analyze policy rules in competitive environments will be needed.
Carnegie-Mellon University
and
University of Minnesota
Manuscript received January, 1981; revision received January, 1982.
19
Examples of such policy issues are described in [21]. See also Barro (e.g., [3]), who
emphasizes the differences in effects of temporary and permanent changes in government
expenditures.
REFERENCES
[1] ABEL, A.B.: “Empirical Investment Equations: An Integrated Framework,” in On The State of Macroeconomics, ed. by
K.Brunner and A.H.Meltzer. Amsterdam: North-Holland, 1979.
[2] ABOWD, J., AND O.ASHENFELTER: “Unemployment and Compensating Wage Differentials,” in Studies in Labor
Markets, ed. by S.Rosen. Chicago: University of Chicago Press, 1981.
[3] BARRO, R.J.: “Intertemporal Substitution and the Business Cycle,” in Supply Shocks, Incentives and National Wealth, ed. by
K.Brunner and A.H.Meltzer. Amsterdan: North-Holland, 1981.
[4] BLACK, F.: “General Equilibrium and Business Cycles,” Working Paper, Massachusetts Institute of Technology, revised
November, 1979.
[5] BLINDER, A.S., AND S.FISCHER: “Inventories, Rational Expectations, and the Business Cycle,” Journal of Monetary
Economics, 8 (1981), 227–304.
[6] BÖHM-BAWERK, E. VON: Positive Theory of Capital, trans. by W.Smart. London: Macmillan, 1891.
[7] BRUNNER, K., AND A.H.MELTZER: “Money, Debt, and Economic Activity,” Journal of Political Economy, 80 (1972),
951–977.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
82
FOUNDATIONS
1370
F.E. KYDLAND AND E.C.PRESCOTT
[8] BRUNNER, K., A. CUKIERMAN, AND A.H.MELTZER: “Stagflation, Persistent Unemployment and the Permanence
of Economic Shocks,” Journal of Monetary Economics, 6(1980), 467–492.
[9] CARLTON, D.W.: “Contracts, Price Rigidity, and Market Equilibrium,” Journal of Political Economy, 87(1979), 1034–1062.
[10] CHRISTIANO, L.J.: “On the Accuracy of Linear Quadratic Approximations: An Example,” Working Paper, University
of Chicago, September, 1981.
[11] DENISON, E.F.: Accounting for United States Economic Growth 1929–1969. Washington, D.C.: Brookings Institution, 1974.
[12] GROSSMAN, H.I.: “Aggregate Demand, Job Search, and Employment,” Journal of Political Economy, 81(1973), 1353–
1369.
[13] HALL, R.E.: “Investment, Interest Rates, and the Effects of Stabilization Policies,” Brookings Papers on Economic Activity,
6(1977), 61–100.
[14] HALL, R.E., AND D.W.JORGENSON: “Tax Policy and Investment Behavior,” American Economic Review, 57(1967),
391–414.
[15] HANSEN, L.P., AND T.J.SARGENT: “Formulating and Estimating Dynamic Linear Rational Expectations Models,”
Journal of Economic Dynamics and Control, 2(1980), 7–46.
[16] HANSEN, L.P., AND K.J.SINGLETON: “Generalized Instrumental Variables Estimation of Nonlinear Rational
Expectations Models,” Econometrica, 50(1982), 1269–1286.
[17] HAYASHI, F.: “Tobin’s Marginal q and Average q: A Neoclassical Interpretation,” Econometrica, 50(1982), 213–224.
[18] HODRICK, R.J., AND E.C.PRESCOTT: “Post-War U.S. Business Cycles: An Empirical Investigation,” Working Paper,
Carnegie-Mellon University, revised November, 1980.
[19] JORGENSON, D.W.: “Anticipations and Investment Behavior,” in The Brookings Quarterly Econometric Model of the United
States, ed. by J.S.Duesenberry et al. Chicago: Rand McNally, 1965.
[20] KYDLAND, F.E.: “Analysis and Policy in Competitive Models of Business Fluctuations,” Working Paper, CarnegieMellon University, revised April, 1981.
[21] KYDLAND, F.E., AND E.C.PRESCOTT: “A Competitive Theory of Fluctuations and the Feasibility and Desirability
of Stabilization Policy,” in Rational Expectations and Economic Policy, ed. by S.Fischer. Chicago: University of Chicago Press,
1980.
[22] LONG, J.B., JR., AND C.I.PLOSSER: “Real Business Cycles,” Working Paper, University of Rochester, November,
1980.
[23] LUCAS, R.E., JR.: “An Equilibrium Model of the Business Cycle,” Journal of Political Economy, 83(1975), 1113–1144.
[24]——: “Understanding Business Cycles,” in Stabilization of the Domestic and International Economy, ed. by K.Brunner and A.H.Meltzer.
New York: North-Holland, 1977.
[25] LUCAS, R.E., JR, AND L.A.RAPPING: “Real Wages, Employment and Inflation,” Journal of Political Economy, 77(1969),
721–754.
[26] MALKIEL, B.G., G.M.VON FURSTENBERG, AND H.S.WATSON: “Expectations, Tobin’s q, and Industry
Investment,” Journal of Finance, 34(1979), 549–561.
[27] MAYER, T.: “Plant and Equipment Lead Times,” Journal of Business, 33(1960), 127–132.
[28] MOOD, A.M., AND F.A.GRAYBILL: Introduction to the Theory of Statistics, 2nd ed. New York: McGraw-Hill, 1963.
[29] MUSSA, M.: “External and Internal Adjustment Costs and the Theory of Aggregate and Firm Investment,” Economica,
44(1977), 163–178.
[30] PRESCOTT, E.C.: “A Note on Dynamic Programming with Unbounded Returns,” Working Paper, University of
Minnesota, 1982.
[31] PRESCOTT, E.C., AND R.MEHRA: “Recursive Competitive Equilibrium: The Case of Homogeneous Households,”
Econometrica, 48(1980), 1365–1379.
[32] SARGENT, T.J.: “Tobin’s q and the Rate of Investment in General Equilibrium,” in On the State of Macroeconomics, ed. by
K.Brunner and A.H.Meltzer. Amsterdam: North-Holland, 1979.
[33]——: Macroeconomic Theory. New York: Academic Press, 1979.
[34] SIMS, C.A.: “Macroeconomics and Reality,” Econometrica, 48(1980), 1–48.
[35] TAYLOR, J.B.: “Aggregate Dynamics and Staggered Contracts,” Journal of Political Economy, 88(1980), 1–23.
[36] TOBIN, J.: “A General Equilibrium Approach to Monetary Theory,” Journal of Money, Credit, and Banking, 1(1969),
15–29.
[37] TOWNSEND, R.M.: “Forecasting the Forecasts of Others,” Working Paper, Carnegie-Mellon University, revised August,
1981.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
CHAPTER 4
83
Edward C.Prescott
Theory Ahead of Measurement
Theory Ahead of Business Cycle Measurement*
Edward C.Prescott
Adviser
Research Department
Federal Reserve Bank of Minneapolis
and Professor of Economics
University of Minnesota
Economists have long been puzzled by the
observations that during peacetime industrial
market economies display recurrent, large
fluctuations in output and employment over
relatively short time periods. Not uncommon
are changes as large as 10 percent within only a
couple of years. These observations are
considered puzzling because the associated
movements in labor’s marginal product are small.
These observations should not be puzzling,
for they are what standard economic theory
predicts. For the United States, in fact, given
people’s ability and willingness to
intertemporally and intratemporally substitute
consumption and leisure and given the nature
of the changing production possibility set, it
would be puzzling if the economy did not display
these large fluctuations in output and
employment with little associated fluctuations
in the marginal product of labor. Moreover,
standard theory also correctly predicts the
amplitude of these fluctuations, their serial
correlation properties, and the fact that the
investment component of output is about six
times as volatile as the consumption component.
This perhaps surprising conclusion is the
principal finding of a research program initiated
by Kydland and me (1982) and extended by
Kydland and me (1984), Hansen (1985a), and
Bain (1985). We have computed the competitive
equilibrium stochastic process for variants of
the constant elasticity, stochastic growth model.
The elasticities of substitution and the share
parameters of the production and utility functions
are restricted to those that generate the growth
observations. The process governing the
technology parameter is selected to be consistent
with the measured technology changes for the
American economy since the Korean War. We
ask whether these artificial economies display
fluctuations with statistical properties similar to
those which the American economy has
displayed in that period. They do.1
I view the growth model as a paradigm for
macro analysis—analogous to the supply and
demand construct of price theory. The elasticities
of substitution and the share parameters of the
growth model are analogous to the price and
income elasticities of price theory. Whether or
not this paradigm dominates, as I expect it will,
is still an open question. But the early results
indicate its power to organize our knowledge.
The finding that when uncertainty in the rate of
technological change is incorporated into the
growth model it displays the business cycle
*This paper was presented at a Carnegie-Rochester
Conference on Public Policy and will appear in a volume of the
conference proceedings. It appears here with the kind
permission of Allan H.Meltzer, editor of that volume. The author
thanks Finn E.Kydland for helpful discussions of the issues
reviewed here, Gary D.Hansen for data series and some
additional results for his growth economy, Lars G.M.Ljungqvist
for expert research assistance, Bruce D.Smith and Allan
H.Meltzer for comments on a preliminary draft, and the
National Science Foundation and the Minneapolis Federal
Reserve Bank for financial support. The views expressed herein
are those of the author alone.
1
Others [Barro (1981) and Long and Plosser (1983), for
example] have argued that these fluctuations are not
inconsistent with competitive theory that abstracts from
monetary factors. Our finding is much stronger: standard
theory predicts that the economy will display the business
cycle phenomena.
9
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
84
FOUNDATIONS
phenomena was both dramatic and
unanticipated. I was sure that the model could
not do this without some features of the payment
and credit technologies.
The models constructed within this
theoretical framework are necessarily highly
abstract. Consequently, they are necessarily
false, and statistical hypothesis testing will reject
them. This does not imply, however, that
nothing can be learned from such quantitative
theoretical exercises. I think much has already
been learned and confidently predict that much
more will be learned as other features of the
environment are introduced. Prime candidates
for study are the effects of public finance
elements, a foreign sector, and, of course,
monetary factors. The research I review here is
best viewed as a very promising beginning of a
much larger research program.
The Business Cycle Phenomena
The use of the expression business cycle is
unfortunate for two reasons. One is that it leads
people to think in terms of a time series’ business
cycle component which is to be explained
independently of a growth component; our
research has, instead, one unifying theory of
both of these. The other reason I do not like to
use the expression is that it is not accurate; some
systems of low-order linear stochastic difference
equations with a nonoscillatory deterministic
part, and therefore no cycle, display key
business cycle features. (See Slutzky 1927.) I
thus do not refer to business cycles, but rather to
business cycle phenomena, which are nothing
more nor less than a certain set of statistical
properties of a certain set of important aggregate
time series. The question I and others have
considered is, Do the stochastic difference
equations that are the equilibrium laws of
motion for the stochastic growth display the
business cycle phenomena?
More specifically, we follow Lucas (1977, p.
9) in defining the business cycle phenomena as
the recurrent fluctuations of output about trend
and the co-movements among other aggregate
time series. Fluctuations are by definition
deviations from some slowly varying path. Since
this slowly varying path increases monotonically
over time, we adopt the common practice of
labeling it trend. This trend is neither a measure
nor an estimate of the unconditional mean of
some stochastic process. It is, rather, defined by
the computational procedure used to fit the
smooth curve through the data.
If the business cycle facts were sensitive to
the detrending procedure employed, there would
be a problem. But the key facts are not sensitive
to the procedure if the trend curve is smooth.
Our curve-fitting method is to take the logarithms
of variables and then select the trend path {τt}
which minimizes the sum of the squared
deviations from a given series {Yt} subject to the
constraint that the sum of the squared second
differences not be too large. This is
The smaller is µ, the smoother is the trend path.
If µ=0, the least squares linear time trend results.
For all series, µ is picked so that the Lagrange
multiplier of the constraint is 1600. This produces
the right degree of smoothness in the fitted trend
when the observation period is a quarter of a
year. Thus, the sequence {τt} minimizes
The first-order conditions of this
minimization problem are linear in Yt and tt,
so for every series, t=AY, where A is the same
T×T matrix. The deviations from trend, also
by definition, are
Unless otherwise stated, these are the variables
used in the computation of the statistics reported
here for both the United States and the growth
economies.
An alternative interpretation of the procedure
is that it is a high pass linear filter. The facts
reported here are essentially the same if, rather
than defining the deviations by Yd=(I-A}Y, we
filtered the Y using a high pass band filter,
eliminating all frequencies of 32 quarters or
greater. An advantage of our procedure is that it
deals better with the ends of the sample problem
and does not require a stationary time series.
To compare the behaviors of a stochastic
growth economy and an actual economy, only
identical statistics for the two economies are used.
By definition, a statistic is a real valued function
of the raw time series. Consequently, if a
comparison is made, say, between the standard
deviations of the deviations, the date t deviation
for the growth economy must be the same
function of the data generated by that model as
the date t deviation for the American economy
is of that economy’s data. Our definitions of the
deviations satisfy this criterion.
10
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THEORY AHEAD OF MEASUREMENT
85
Edward C.Prescott
Theory Ahead of Measurement
Figure 1 plots the logs of actual and trend
output for the U.S. economy during 1947–82,
and Figure 2 the corresponding percentage
deviations from trend of output and hours of
market employment. Output and hours clearly
move up and down together with nearly the
same amplitudes.
Table 1 contains the standard deviations and
cross serial correlations of output and other
aggregate time series for the American economy
during 1954–82. Consumption appears less
variable and investment more variable than
output. Further, the average product of labor is
procyclical but does not vary as much as output
or hours.
Figure 1
Actual and Trend Logs of U.S. Gross National Product
Quarterly, 1947–82
The Growth Model
This theory and its variants build on the
neoclassical growth economy of Solow (1956)
and Swan (1956). In the language of Lucas
(1980, p. 696), the model is a “fully articulated,
artificial economic system” that can be used to
generate economic time series of a set of important
econom c aggregates. The model assumes an
aggregate production function with constant
returns to scale, inputs labor n and capital k, and
an output which can be allocated either to current
consumption c or to investment x. If t denotes
the date, f: R2→R the production function, and zt
a technology parameter, then the production
constraint is
Source of basic data: Citicorp’s Citibase data bank
Figure 2
Deviations From Trend of Gross National Product
and Nonfarm Employee Hours in the United States
Quarterly, 1947–82
xt+ctⱕztf(kt, nt)
where xt, ct, kt, nt ⱖ0. The model further assumes
that the services provided by a unit of capital
decrease geometrically at a rate 0<δ<1:
kt+1=(1-δ)kt+xt.
Solow completes the specification of his
economy by hypothesizing that some fraction
0<σ<1 of output is invested and the remaining
fraction 1-σ consumed and that nt is a constant—
say, n-—for all t. For this economy, the law of
motion of capital condition on zt is
Once the {zt} stochastic process is specified, the
stochastic process governing capital and the other
economic aggregates are determined and
realizations of the stochastic process can be
generated by a computer.
This structure is far from adequate for the study of
Source of basic data: Citicorp’s Citibase data bank
the business cycle because in it neither
employment nor the savings rate varies, when in
fact they do. Being explicit about the economy,
however, naturally leads to the question of what
determines these variables, which are central to
the cycle.
That leads to the introduction of a stand-in
household with some explicit preferences. If we
11
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
86
FOUNDATIONS
Table 1
Cyclical Behavior of the U.S. Economy
Deviations From Trend of Key Variables, 1954:1–1982:4
Source of basic data: Citicorp's Citibase data bank
abstract from the labor supply decision and
and
), the
uncertainty (that is,
standard form of the utility function is
where β is the subjective time discount factor.
The function u:R+→R is twice differentiable and
concave. The commodity space for the
deterministic version of this model is l∞, infinite
sequences of uniformly bounded consumptions
The theorems of Bewley (1972) could be
applied to establish existence of a
competitive equilibrium for this l ∞
commodity-space economy. That existence
argument, however, does not provide an
algorithm for computing the equilibria. An
alternative approach is to use the competitive
welfare theorems of Debreu (1954). Given
local nonsaturation and no externalities,
competitive equilibria are Pareto optima and,
with some additional conditions that are
satisfied for this economy, any Pareto
optimum can be supported as a competitive
equilibrium. Given a single agent and the
convexity, there is a unique optimum and
that optimum is the unique competitive
equilibrium allocation. The advantage of this
approach is that algorithms for computing
solutions to concave programming problems
can be used to find the competitive
equilibrium allocation for this economy.
Even with the savings decision endogenous,
this economy has no fluctuations. As shown by
Cass (1965) and Koopmans (1965), the
competitive equilibrium path converges
monotonically to a unique rest point or, if zt is
growing exponentially, to a balanced growth
path. There are multisector variants of this model
12
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THEORY AHEAD OF MEASUREMENT
87
Edward C.Prescott
Theory Ahead of Measurement
in which the equilibrium path oscillates. (See
Benhabib and Nishimura 1985 and Marimon
1984.) But I know of no multisector model which
has been restricted to match observed factor
shares by sector, which has a value for β
consistent with observed interest rates, and which
displays oscillations.
When uncertainty is introduced, the
household’s objective is its expected discounted
utility:
The commodity vector is now indexed by the
is
history of shocks; that is,
the commodity point. As Brock and Mirman
(1972) show, if the {zt} are identically distributed
random variables, an optimum to the social
planner’s problem exists and the optimum is a
stationary stochastic process with kt+1=g(kt, zt) and
ct=c(kt, zt). As Lucas and Prescott (1971) show, for
a class of economies that include this one, the
social optimum is the unique competitive
equilibrium allocation. They also show that for
these homogeneous agent economies, the social
optimum is also the unique sequence-of-markets
equilibrium allocation. Consequently, there are
equilibrium time-invariant functions for the wage
wt=w(kt, zt) and the rental price of capital rt=r(kt,
zt), where these prices are relative to the date t
consumption good. Given these prices, the firm’s
period t problem is
subject to the output constraint
ytⱕztf(kt, nt).
The household’s problem is more
complicated, for it must form expectations of future
prices. If at is its capital stock, its problem is and
given a0-k0. In forming expectations, a household
knows the relation between the economy’s state
(kt, zt) and prices, wt=w(kt, zt) and rt=r(kt, zt).
Further, it knows the process governing
the evolution of the per capita capital stock, a
variable which, like prices, is taken as given.
The elements needed to define a sequence-ofmarkets equilibrium are the firm’s policy functions
y(kt, zt), n(kt, zt), and k(kt, zt); the household’s policy
functions x(at, kt, zt) and c(at, kt, zt); a law of motion
of per capita capital kt+1=g(kt, zt); and pricing
functions w(kt, zt) and r(kt, zt). For equilibrium,
then,
•
The firm’s policy functions must be optimal given the pricing functions.
•
The household’s policy functions must be
optimal given the pricing functions and the
law of motion of per capita capital.
•
Spot markets clear; that is, for all kt and zt
(Note that the goods market must clear only
when the representative household is truly
representative, that is, when at=kt.)
•
Expectations are rational; that is,
g(kt, zt)=(1-δ)kt+x(kt, kt, zt).
This definition still holds if the household values
productive time that is allocated to nonmarket
activities. Such time will be called leisure and
denoted lt. The productive time endowment is
normalized to 1, and the household faces the
constraints
n t +l t ⱕ 1
for all t. In addition, leisure is introduced as an
argument of the utility function, so the
household’s objective becomes the
maximization of
Now leisure—and therefore employment—varies
in equilibrium.
The model needs one more modification: a
relaxation of the assumption that the technology
shocks zt are identically and independently
distributed random variables. As will be
documented, they are not so distributed. Rather,
13
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
88
FOUNDATIONS
they display considerable serial correlation, with
their first differences nearly serially uncorrelated.
To introduce high persistence, we assume
zt+1=␳zt+⑀t+1
where the {⑀t+1} are identically and independently
distributed and ␳ is near 1. With this
modification, the recursive sequence-of-markets
equilibrium definition continues to apply.
Using Data to Restrict the Growth Model
Without additional restrictions on preferences
and technology, a wide variety of equilibrium
processes are consistent with the growth
model. The beauty of this model is that both
growth and micro observations can be used to
determine its production and utility functions.
When they are so used, there are not many
free parameters that are specific to explaining
the business cycle phenomena and that cannot
be measured independently of those
phenomena. The key parameters of the growth
model are the intertemporal and intratemporal
elasticities of substitution. As Lucas (1980, p.
712) emphasizes, “On these parameters, we
have a wealth of inexpensively available data
from census cohort information, from panel
data describing the reactions of individual
households to a variety of changing market
conditions, and so forth.” To this list we add
the secular growth observations which have
the advantage of being experiments run by
nature with large changes in relative prices and
quantities and with idiosyncratic factors
averaged out.2 A fundamental thesis of this line
of inquiry is that the measures obtained from
aggregate series and those from individual
panel data must be consistent. After all, the
former are just the aggregates of the latter.
Secularly in the United States, capital and
labor shares of output have been approximately
constant, as has r, the rental price of capital.
However, the nation’s real wage has increased
greatly—more than 100 percent since the Korean
War. For these results to hold, the model’s
production function must be approximately
Cobb-Douglas:
The share parameter ␪ is equal to labor’s share,
which has been about 64 percent in the postwar
period, so ␪=0.64. This number is smaller than
that usually obtained because we include services
of consumer durables as part of output. This
alternative accounting both reduces labor’s share
and makes it more nearly constant over the
postwar period.
The artificial economy has but one type of
capital, and it depreciates at rate ␦. In fact, different
types of capital depreciate at different rates, and
the pattern of depreciation over the life of any
physical asset is not constant. Kydland and I (1982,
1984) simply pick ␦=0.10. With this value and
an annual real interest rate of 4 percent, the steadystate capital-annual output ratio is about 2.6. That
matches the ratio for the U.S economy and also
implies a steady-state investment share of output
near the historically observed average Except for
parameters determining the process on the
technology shock, this completely specifies the
technology of the simple growth model.
A key growth observation which restricts the
utility function is that leisure per capita lt has
shown virtually no secular trend while, again,
the real wage has increased steadily. This
implies an elasticity of substitution between
consumption ct and leisure lt near 1. Thus the
utility function restricted to display both
constant intertemporal and unit intratemporal
elasticities of substitution is
where 1/␥>0 is the elasticity of substituting
between different date composite commodities
This leaves ␥ and the subjective time
discount factor β [or, equivalently, the subjective
time discount rate (1/β)-1] to be determined.
The steady-state interest rate is
i=(1/β)-1+␥(c/c).
As stated previously, the average annual real
interest rate is about 4 percent, and the growth
rate of per capita consumption c/c has averaged
nearly 2 percent. The following studies help
restrict γ. Tobin and Dolde (1971) find that a γ
near 1.5 is needed to match the life cycle
consumption patterns of individuals. Using
individual portfolio observations, Friend and
Blume (1975) estimate γ to be near 2. Using
aggregate stock market and consumption data,
Hansen and Singleton (1983) estimate γ to be
near 1. Using international data, Kehoe (1984)
also finds a modest curvature parameter ␥. All
these observations make a strong case that ␥ is
not too far from 1. Since the nature of
fluctuations of the artificial economy is not very
2
See Solow 1970 for a nice summary of the growth
observations.
14
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THEORY AHEAD OF MEASUREMENT
89
Edward C.Prescott
Theory Ahead of Measurement
sensitive to ␥, we simply set ␥ equal to 1. Taking
the limit as ␥→1 yields
u(ct, lt)=(1-␾) log ct+ ␾log lt.
This leaves β and ␾ still to be determined.
Hansen (1985b) has found that growing
economies—that is, those with z t having a
multiplicative, geometrically growing factor
(1+λ)t with λ>0—fluctuate in essentially the same
way as economies for which λ=0. This justifies
considering only the case λ=0. If λ=0, then the
average interest rate approximately equals the
subjective time discount rate.3 Therefore, we set
β equal to 0.96 per year or 0.99 per quarter.
The parameter ␾ is the leisure share parameter.
Ghez and Becker (1975) find that the household
allocates approximately one-third of its productive
time to market activities and two-thirds to nonmarket
activities. To be consistent with that, the model’s
parameter ␾ must be near two-thirds. This is the
value assumed in our business cycle studies.
Eichenbaum, Hansen, and Singleton (1984)
use aggregate data to estimate this share parameter
␾, and they obtain a value near five-sixths. The
difference between two-thirds and five-sixths is large
in the business cycle context. With ␾=2/3, the
elasticity of labor supply with respect to a temporary
change in the real wage is 2, while if ␾=5/6, it is 5.
This is because a 1 percent change in leisure implies
a ␾/(␾-1) percent change in hours of employment.
We do not follow the Eichenbaum-HansenSingleton approach and treat ␾ as a free
parameter because it would violate the principle
that parameters cannot be specific to the
phenomena being studied. What sort of science
would economics be if micro studies used one
share parameter and aggregate studies another?
The Nature of the Technological Change
One method of measuring technological change
is to follow Solow (1957) and define it as the
changes in output less the sum of the changes in
labor’s input times labor share and the changes in
capital’s input times capital share. Measuring
variables in logs, this is the percentage change in
the technology parameter of the Cobb-Douglas
production function. For the U.S. economy
between the third quarter of 1955 and the first
quarter of 1984, the standard deviation of this
change is 1.2 percent.4 The serial autocorrelations
of these changes are ␳1=-0.21, ␳2=-0.06, ␳3=0.04,
␳4= 0.01, and ␳5=-0.05. To a first approximation,
the process on the percentage change in the
technology process is a random walk with drift
plus some serially uncorrelated measurement
error. This error produces the negative first-order
serial correlation of the differences.
Further evidence that the random walk model
is not a bad approximation is based on yearly
changes. For the quarterly random walk model,
the standard deviation of this change is 6.63 times
the standard deviation of the quarterly change.
For the U.S. data, the annual change is only 5.64
times as large as the quarterly change. This, along
with the negative first-order serial correlation,
suggests that the standard deviation of the
persistent part of the quarterly change is closer to
5.64/6.63= 0.85 than to 1.2 percent. Some further
evidence is the change over four-quarter periods—
that is, the change from a given quarter of one
year to the same quarter of the next year. For the
random walk model, the standard deviation of
these changes is 2 times the standard deviation of
the quarterly change. A reason that the standard
deviation of change might be better measured
this way is that the measurement noise introduced
by seasonal factors is minimized. The estimate
obtained in this way is 0.95 percent. To
summarize, Solow growth accounting finds that
the process on the technology parameter is highly
persistent with the standard deviation of change
being about 0.90.5
The Solow estimate of the standard
deviation of technological change is surely an
overstatement of the variability of that
parameter. There undoubtedly are nonnegligible errors in measuring the inputs.
Since the capital input varies slowly and its
share is small, the most serious measurement
problem is with the labor input. Fortunately
there are two independent measures of the
aggregate labor input, one constructed from a
survey of employers and the other from a
survey of households. Under the assumption
of orthogonality of their measurement errors,
a reasonable estimate of the variance of the
change in hours is the covariance between the
changes in the two series. Since the
household survey is not used to estimate
aggregate output, I use the covariance
between the changes in household hours and
output as an estimate of the covariance
3
Actually, the average interest rate is slightly lower because
of risk premia. Given the value of ␥ and the amount of
uncertainty, the average premium is only a fraction of a
percent. See Mehra and Prescott 1985 for further details.
4
I use Hansen’s (1984) human capital-weighted,
household hour series. The capital stock and GNP series are
from Citicorp’s Citibase data bank.
5
The process zt+1=.9zt+et+1 is, like the random walk
process, highly persistent. Kydland and I find that it and the
random walk result in essentially the same fluctuations.
15
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
90
FOUNDATIONS
between aggregate hours and output. Still
using a share parameter of ␪ =0.75, my
estimate of the standard deviation of the
percentage change in zt is the square root of
ˆ
␪ cov(Δhˆ 1 , Δŷ)+ ␪ 2 cov(Δhˆ 1, Δhˆ 2),
var(Δy)-2
where the caret ( ˆ ) denotes a measured value.
For the sample period my estimate is 0.763
percent. This is probably a better estimate
than the one which ignores measurement
error.
Still, my estimate might under- or overstate
the variance of technological change. For
example, the measurement of output might
include significant errors. Perhaps
measurement procedures result in some
smoothing of the series. This would reduce
the variability of the change in output and
might reduce the covariance between
measured hours and output.
Another possibility is that changes in hours
are associated with corresponding changes in
capital’s utilization rate. If so, the Solow approach
is inappropriate for measuring the technology
shocks. To check whether this is a problem, I
varied ␪ and found that ␪= 0.85 yields the smallest
estimate, 0.759, as opposed to 0.763 for ␪=0.75.
This suggests that my estimate is not at all sensitive
to variations in capital utilization rates.
To summarize, there is overwhelming
evidence that technological shocks are highly
persistent. But tying down the standard
deviation of the technology change shocks is
difficult. I estimate it as 0.763. It could very
well be larger or smaller, though, given the
accuracy of the measurements.
The Statistical Behavior of the Growth
Models
Theory provides an equilibrium stochastic
process for the growth economy studied. Our
approach has been to document the similarities
and differences between the statistical properties
of data generated by this stochastic process and
the statistical properties of American time series
data. An alternative approach is to compare the
paths of the growth model if the technological
parameters {zt} were those experienced by the
U.S. economy. We did not attempt this because
theory’s predictions of paths, unlike its
predictions of the statistical properties, are
sensitive to what Learner (1983, p. 43) calls
“whimsical” modeling assumptions. Another
nontrivial problem is that the errors in measuring
the innovations in the zt process are as large as
the innovations themselves.
The Basic Growth Model
With the standard deviation of the technology
shock equal to 0.763, theory implies that the
standard deviation of output will be 1.48
percent. In fact, it is 1.76 percent for the postKorean War American economy. For the output
of the artificial economy to be as variable as that,
the variance of the shock must be 1.0,
significantly larger than the estimate. The most
important deviation from theory is the relative
volatility of hours and output. Figure 3 plots a
realization of the output and employment
deviations from trend for the basic growth
economy. A comparison of Figures 2 and 3
demonstrates clearly that, for the American
economy, hours in fact vary much more than
the basic growth model predicts. For the
artificial economy, hours fluctuate 52 percent as
much as output, whereas for the American
economy, the ratio is 0.95. This difference
appears too large to be a result of errors in
measuring aggregate hours and output.
The Kydland-Prescott Economy
Kydland and I (1982, 1984) have modified the
growth model in two important respects. First,
we assume that a distributed lag of leisure and
the market-produced good combine to produce
the composite commodity good valued by the
household. In particular,
where ␣i+1/␣i=1-␩ for i=1, 2, …and
Figure 3
Deviations From Trend of GNP and Hours Worked
in the Basic Growth Economy
16
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THEORY AHEAD OF MEASUREMENT
91
Edward C.Prescott
Theory Ahead of Measurement
Table 2
Cyclical Behavior of the Kydland-Prescott Economy*
* These are the means of 20 simulations, each of which was 116 periods long. The numbers
in parentheses are standard errors. Source: Kydland and Prescott 1984
Kydland (1983) provides justification for this
preference ordering based on an unmeasured,
household-specific capital stock that, like ct and
lt, is an input in the production of the composite
commodity. The economy studied has ␣0=0.5
and ␩=0.1. This increases the variability of hours.
The second modification is to permit the
workweek of capital to vary proportionally to
the workweek of the household. For this
economy, increases in hours do not reduce the
marginal product of labor as much, so hours
fluctuate more in response to technology shocks
of a given size.
The statistical properties of the fluctuations
for this economy are reported in Table 2. As is
clear there, hours are now about 70 percent as
variable as output. This eliminates much of the
discrepancy between theory and measurement.
If the standard deviation of the technology shock
is 0.72 percent, then fluctuations in the output of
this artificial economy are as large as those
experienced in the U.S. economy.
A comparison of Tables 1 and 2 shows that
the Kydland-Prescott economy displays the
business cycle phenomena. It does not quite
demonstrate, however, that there would be a
puzzle if the economy did not display the business
cycle phenomena. That is because the parameters
␣0 and ␩ have not been well tied down by micro
observations.6 Better measures of these parameters
could either increase or decrease significantly
the amount of the fluctuations accounted for by
the uncertainty in the technological change.
6
Hotz, Kydland, and Sedlacek (1985) use annual panel
data to estimate ␣0 and ␩ and obtain estimates near the
Kydland-Prescott assumed values.
17
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
92
FOUNDATIONS
The Hansen Indivisible Labor Economy
Labor economists have estimated labor supply
elasticities and found them to be small for fulltime prime-age males. (See, for example,
Ashenfelter 1984.) Heckman (1984), however,
finds that when movements between
employment and nonemployment are
considered and secondary workers are
included, elasticities of labor supply are much
larger. He also finds that most of the variation in
aggregate hours arises from variation in the
number employed rather than in the hours
worked per employed person.
These are the observations that led Hansen
(1985a) to explore the implication of
introducing labor indivisibilities into the
growth model. As shown by Rogerson (1984),
if the household’s consumption possibility set
has nonconvexities associated with the
mapping from hours of market production
activities to units of labor services, there will be
variations in the number employed rather than
in the hours of work per employed person. In
addition, the aggregate elasticity of labor
supply will be much larger than the elasticity of
those whose behavior is being aggregated. In
this case aggregation matters, and matters
greatly.
There
certainly
are
important
nonconvexities in the mapping from hours of
market activities to units of labor services
provided. Probably the most important
nonconvexity arises from the considerable
amount of time required for commuting. Other
features of the environment that would make
full-time workers more than twice as productive
as otherwise similar half-time workers are not
hard to imagine. The fact that part-time workers
typically are paid less per hour than full-time
workers with similar human capital
endowments is consistent with the existence of
important nonconvexities.
Hansen (1985a) restricts -each identical
household to either work h hours or be
unemployed. His relation is as depicted by the
horizontal lines in Figure 4. This assumption is
not as extreme as it appears. If the relation were
as depicted by the curved line, the behavior of
the economy would be the same. The key
property is an initial convex region followed by
a concave region in the mapping from hours of
market activity to units of labor service.
Figure 4
Relation Between Time Allocated to Market Activity
and Labor Service
With this modification, lotteries that specify
the probability of employment are traded along
with market-produced goods and capital
services. As before, the utility function of each
individual is
u(c, l)=(1/3) log c+(2/3) log l.
otherwise,
If an individual works,
l=1. Consequently, if ␲ is the probability of
employment, an individual’s expected utility is
Given that per capita consumption is c- and
per capita hours of employment n-, average
utility over the population is maximized
by
for all individuals. If l , which equals
setting
1–␲h , denotes per capita leisure, then maximum
per capita utility is
This is the utility function which rationalizes the
per capita consumption and leisure choices if
each person’s leisure is constrained to be either
or 1. The aggregate intertemporal elasticity
of substitution between different date leisures is
infinity independent of the value of the elasticity
for the individual (in the range where not all are
employed).
Hansen (1985a) finds that if the technology
shock standard deviation is 0.71, then fluctuations
in output for his economy are as large as those
for the American economy. Further, variability
18
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THEORY AHEAD OF MEASUREMENT
93
Edward C.Prescott
Theory Ahead of Measurement
in hours is 77 percent as large as variability in
output. Figure 5 shows that aggregate hours and
output for his economy fluctuate together with
nearly the same amplitude. These theoretical
findings are the basis for my statement in the
introduction that there would be a puzzle if the
economy did not display the business cycle
phenomena.
Empirical Labor Elasticity
One important empirical implication of a shockto-technology theory of fluctuations is that the
empirical labor elasticity of output is significantly
larger than the true elasticity, which for the CobbDouglas production function is the labor share
parameter. To see why, note that the capital stock
varies little cyclically and is nearly uncorrelated
with output. Consequently, the deviations almost
satisfy
yt=␪ht+zt
where yt is output, ht hours, and zt the technology
shock. The empirical elasticity is
␩=cov(ht, yt)/var(ht)
which, because of the positive correlation between
ht and zt, is considerably larger than the model’s
␪, which is 0.64. For the basic, Kydland-Prescott,
and Hansen growth economies, the values of ␩
are 1.9, 1.4, and 1.3, respectively.
Because of measurement errors, the empirical
elasticity for the American economy is not wellestimated by simply computing the ratio of the
covariance between hours and output and
dividing by the variance of hours. The procedure
I use is based on the following probability model:
where the caret ( ˆ ) denotes a measured value.
The ⑀it are measurement errors. Here, the hˆ1t
measure of hours uses the employer survey data
while the hˆ2t measure uses the household survey
data. Since these are independent measures, a
maintained hypothesis is that ⑀2t and ⑀3t are
orthogonal. With this assumption, a reasonable
Figure 5
Deviations From Trend of GNP and Hours Worked in
Hansen’s Indivisible Labor Economy
Source Gary D Hansen. Department of Economics. University
of California Santa Barbara
estimate of var(ht) is the sample covariance
between hˆ1t and hˆ2t. Insofar as the measurement
of output has small variance or ⑀1t is uncorrelated
with the hours measurement errors or both, the
covariance between measured output and either
measured hours series is a reasonable estimate of
the covariance between output and hours. These
two covariances are 2.231×10-4 and 2.244×10-4
for the sample period, and I take the average as
my estimate of cov(ht, yt) for the American
economy. My estimate of the empirical labor
elasticity of output is
This number is considerably greater than labor’s
share, which is about 0.70 when services of
consumer durables are not included as part of
output. This number strongly supports the
importance of technological shocks in accounting
for business cycle fluctuations. Nevertheless, the
number is smaller than those for the KydlandPrescott and Hansen growth economies.
One possible reason for the difference
between the U.S. economy and the growth
model empirical labor elasticities of output is
cyclical measurement errors in output. A sizable
part of the investment component of output is
hard to measure and therefore not included in
19
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
94
FOUNDATIONS
the U.S. National Product Accounts measure of
output, the gross national product (GNP). In
particular, a firm’s major maintenance
expenditures, research and development
expenditures, and investments in human capital
are not included in GNP. In good times—namely,
when output is above trend—firms may be more
likely to undertake major repairs of a not fully
depreciated asset, such as replacing the roof of a
30-year-old building which has a tax life of 35
years. Such an expenditure is counted as
maintenance and therefore not included in GNP
even though the new roof will provide productive
services for many years. The incentive for firms
to do this is tax savings: by expensing an
investment rather than capitalizing it, current tax
liabilities are reduced. Before 1984, when a
railroad replaced its 90-pound rails, the
expenditure was treated as a maintenance
expense rather than an investment expenditure.
If these and other types of unmeasured investment
fluctuate in percentage terms more than output,
as do all the measured investment components,
the volatility of GNP is larger than measured.
We do know that investment in rails was highly
procyclical and volatile in the postwar period. A
careful study is needed to determine whether the
correction for currently unmeasured investment
is small or large.
Another reason to expect the American
economy’s labor elasticity to be less than the model’s
is that the model shocks are perfectly neutral with
respect to the consumption and investment good
transformation. Persistent shocks which alter the
product transformation frontier between these goods
would cause variation in output and employment
but not in the productivity parameters. For
fluctuations so induced, the empirical labor
elasticity of output would be the true elasticity.
Similarly, relatively permanent changes in the taxing
of capital—such as altering depreciation rates, the
corporate income tax rate, or the investment tax
credit rate—would all result in fluctuations in output
and employment but not in the productivity
parameters.
A final reason for actual labor elasticity to be
less than the model’s is the way imports are
measured. An increase in the price of imported
oil, that is, an increase in the quantity of output
that must be sacrificed for a given unit of that
input, has no effect on measured productivity.
From the point of view of the growth model,
however, an oil price increase is a negative
technology shock because it results in less output,
net of the exports used to pay for the imported
oil, available for domestic consumption and
investment. Theory predicts that such shocks
will induce variations in employment and output,
even though they have no effect on the aggregate
production function. Therefore, insofar as they
are important, they reduce the empirical labor
elasticity of output.
Extensions
The growth model has been extended to provide
a better representation of the technology. Kydland
and I (1982) have introduced a technology with
more than one construction period for new
production capacity.7 We have also introduced
inventory as a factor of production. This
improves the match between the model’s serial
correlation properties and the U.S. postwar data,
but has little effect on the other statistics.
Kydland (1984) has introduced heterogeneity
of labor and found that if there are transfers from
high human capital people to low human capital
people, theory implies that hours of the low
fluctuate more than hours of the high. It also
implies a lower empirical labor elasticity of
output than the homogeneous household model.
Bain (1985) has studied an economy that is
richer in sectoral detail. His model has
manufacturing, retailing, and service-producing
sectors. A key feature of the technology is that
production and distribution occur sequentially.
Thus there are two types of inventories—those of
manufacturers’ finished goods and those of final
goods available for sale. With this richer detail,
theory implies that different components of
aggregate inventories behave in different ways,
as seen in the data. It also implies that production
is more volatile than final sales, an observation
considered anomalous since inventories can be
used to smooth production. (See, for example,
Blinder 1984.)
Much has been done. But much more remains
to be explored. For example, public finance
considerations could be introduced and theory
used to predict their implications. As mentioned
above, factors which affect the rental price of capital
affect employment and output, and the nature of
the tax system affects the rental price of capital.
Theory could be used to predict the effect of
temporary increases in government expenditures
such as those in the early 1950s when defense
expenditures increased from less than 5 to more
than 13 percent of GNP. Theory of this type
could also be used to predict the effect of termsof-trade shocks. An implication of such an
exercise most likely will be that economies with
persistent terms-of-trade shocks fluctuate
7
Altug (1983) has introduced two types of capital with
different gestation periods. Using formal econometric
methods, she finds evidence that the model’s fit is improved
if plant and equipment investment are not aggregated.
20
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THEORY AHEAD OF MEASUREMENT
95
Edward C.Prescott
Theory Ahead of Measurement
differently than economies with transitory
shocks. If so, this prediction can be tested against
the observations.
Another interesting extension would be to
explicitly model household production. This
production often involves two people, with one
specializing in market production and the other
specializing in household production while
having intermittent or part-time market
employment. The fact that, cyclically, the
employment of secondary wage earners is much
more volatile than that of primary wage earners
might be explained.
A final example of an interesting and not yet
answered question is, How would the behavior
of the Hansen indivisible labor economy change
if agents did not have access to a technology to
insure against random unemployment and
instead had to self-insure against unemployment
by holding liquid assets? In such an economy,
unlike Hansen’s, people would not be happy
when unemployed. Their gain of more leisure
would be more than offset by their loss as an
insurer. Answering this question is not
straightforward, because new tools for computing
equilibria are needed.
Summary and Policy Implications
Economic theory implies that, given the nature
of the shocks to technology and people’s
willingness and ability to intertemporally and
intratemporally substitute, the economy will
display fluctuations like those the U.S. economy
displays. Theory predicts fluctuations in output
of 5 percent and more from trend, with most of
the fluctuation accounted for by variations in
employment and virtually all the rest by the
stochastic technology parameter. Theory predicts
investment will be three or more times as volatile
as output and consumption half as volatile.
Theory predicts that deviations will display high
serial correlation. In other words, theory predicts
what is observed. Indeed, if the economy did
not display the business cycle phenomena, there
would be a puzzle.
The match between theory and observation
is excellent, but far from perfect. The key
deviation is that the empirical labor elasticity of
output is less than predicted by theory. An
important part of this deviation could very well
disappear if the economic variables were
measured more in conformity with theory. That
is why I argue that theory is now ahead of
business cycle measurement and theory should
be used to obtain better measures of the key
economic time series. Even with better
measurement, there will likely be significant
deviations from theory which can direct
subsequent theoretical research. This feedback
between theory and measurement is the way
mature, quantitative sciences advance.
The policy implication of this research is
that costly efforts at stabilization are likely to be
counterproductive. Economic fluctuations are
optimal responses to uncertainty in the rate of
technological change. However, this does not
imply that the amount of technological change
is optimal or invariant to policy. The average
rate of technological change varies much both
over time within a country and across national
economies. What is needed is an understanding
of the factors that determine the average rate at
which technology advances. Such a theory
surely will depend on the institutional
arrangements societies adopt. If policies adopted
to stabilize the economy reduce the average rate
of technological change, then stabilization policy
is costly. To summarize, attention should be
focused not on fluctuations in output but rather
on determinants of the average rate of
technological advance.
21
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
96
FOUNDATIONS
References
Altug, S. 1983. Gestation lags and the business cycle. Working
paper, Carnegie-Mellon University.
Ashenfelter, O. 1984. Macroeconomic analyses and
microeconomic analyses of labor supply. In Essays on
macroeconomic implications of financial and labor markets and
political processes, ed. K.Brunner and A.H.Meltzer.
Carnegie-Rochester Conference Series on Public Policy
21:117–55. Amsterdam: North-Holland.
Bain, I.R.M. 1985. A theory of the cyclical movements of
inventory stocks. Ph.D. dissertation, University of
Minnesota.
Barro, R.J. 1981. Intertemporal substitution and the business
cycle. In Supply shocks, incentives and national wealth, ed.
K.Brunner and A.H.Meltzer. Carnegie-Rochester
Conference Series on Public Policy 14:237–68.
Amsterdam: North-Holland.
Benhabib, J., and Nishimura, K. 1985. Competitive
equilibrium cycles. Journal of Economic Theory 35:284–
306.
Bewley, T.F. 1972. Existence of equilibria in economies with
infinitely many commodities. Journal of Economic Theory
4:514–40.
Blinder, A.S. 1984. Can the production smoothing model of
inventory behavior be saved? Working paper,
Princeton University.
Brock, W.A., and Mirman, L.J. 1972. Optimal economic
growth and uncertainty: The discounted case. Journal
of Economic Theory 4: 479–513.
Cass, D. 1965. Optimum growth in an aggregative model of capital
accumulation. Review of Economic Studies 32:233–40.
Debreu, G. 1954. Valuation equilibrium and Pareto optimum.
Proceedings of the National Academy of Science 70:558–92.
Eichenbaum, M.S.; Hansen, L.P.; and Singleton, K.S. 1984.
A time series analysis of representative agent models of
consumption and leisure choice under uncertainty.
Working paper, Carnegie-Mellon University.
Friend, I., and Blume, M.E. 1975. The demand for risky assets.
American Economic Review 65:900–22.
Ghez, G.R., and Becker, G.S. 1975. The allocation of time
and goods over the life cycle. New York: National
Bureau of Economic Research.
Hansen, G.D. 1984. Fluctuations in total hours worked: A
study using efficiency units. Working paper, University
of Minnesota.
——. 1985a. Indivisible labor and the business cycle. Journal of
Monetary Economics 16:309–27.
——. 1985b. Growth and fluctuations. Working paper,
University of California, Santa Barbara.
Hansen, L.P., and Singleton, K.J. 1983. Stochastic
consumption, risk aversion, and the temporal behavior
of asset returns. Journal of Political Economy 91:249–65.
Heckman, J. 1984. Comments on the Ashenfelter and
Kydland papers. In Essays on macroeconomic implications of
financial and labor markets and political processes, ed.
K.Brunner and A.H.Meltzer. Carnegie-Rochester
Conference Series on Public Policy 21:209–24.
Amsterdam: North-Holland.
Hotz, V.S.; Kydland, F.E.; and Sedlacek, G.L. 1985.
Intertemporal preferences and labor supply. Working
paper, Carnegie-Mellon University.
Kehoe, P.J. 1984. Dynamics of the current account:
Theoretical and empirical analysis. Working paper,
Harvard University.
Koopmans, T.C. 1965. On the concept of optimal economic
growth. In The econometric approach to development planning.
Chicago: Rand-McNally.
Kydland, F.E. 1983. Nonseparable utility and labor supply.
Working paper, Hoover Institution.
——. 1984. Labor-force heterogeneity and the business cycle.
In Essays on macroeconomic implications of financial and labor
markets and political processes, ed. K.Brunner and
A.H.Meltzer. Carnegie-Rochester Conference Series
on Public Policy 21:173–208. Amsterdam: NorthHolland.
Kydland, F.E., and Prescott, E.C. 1982. Time to build and
aggregate fluctuations. Econometrica 50–70.
——. 1984. The workweek of capital and labor. Research
Department Working Paper 267, Federal Reserve Bank
of Minneapolis.
Leamer, E.E. 1983. Let’s take the con out of econometrics.
American Economic Review 73:31–43.
Long, J.B., and Plosser, C.I. 1983. Real business cycles. Journal
of Political Economy 91:39–69.
Lucas, R.E., Jr. 1977. Understanding business cycles. In
Stabilization of the domestic and international economy, ed.
K.Brunner and A.H.Meltzer. Carnegie-Rochester
Conference Series on Public Policy 5:7–29.
Amsterdam: North-Holland.
——. 1980. Methods and problems in business cycle theory.
Journal of Money, Credit and Banking 12 : 696–715.
Reprinted in Studies in business-cycle theory, pp. 271–96.
Cambridge, Mass.: MIT Press, 1981.
Lucas, R.E., Jr., and Prescott, E.C. 1971. Investment under
uncertainty. Econometrica 39:659–81.
Marimon, R. 1984. General equilibrium and growth under
uncertainty: The turnpike property. Discussion Paper
624. Northwestern University, Center for
Mathematical Studies in Economics and Management
Science.
Mehra, R. and Prescott, E.C. 1985. The equity premium: A
puzzle. Journal of Monetary Economics 15:145–61.
Rogerson, R.D. 1984. Indivisible labor, lotteries and
equilibrium. In Topics in the theory of labor markets, chap. 1.
Ph.D. dissertation, University of Minnesota.
Slutzky, E. 1927. The summation of random causes as the
source of cyclic processes. In Problems of Economic
Conditions, ed. Conjuncture Institute, Moskva
(Moskow), vol. 3, no. 1. Revised English version, 1937,
in Econometrica 5:105–46.
Solow, R.M. 1956. A contribution to the theory of economic
growth. Quarterly Journal of Economics 70:65–94.
——. 1970. Growth theory. Oxford: Oxford University Press.
Swan, T.W. 1956. Economic growth and capital
accumulation. Economic Record 32:334–61.
Tobin, J., and Dolde, W. 1971. Wealth, liquidity and
consumption. In Consumer spending and monetary policy:
The linkages. Monetary Conference Series 5:99–146.
Boston: Federal Reserve Bank of Boston.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
CHAPTER 5
97
Federal Reserve Bank of Minneapolis
Quarterly Review Fall 1986
Some Skeptical Observations
on Real Business Cycle Theory*
Lawrence H.Summers
Professor of Economics
Harvard University and Research Associate
National Bureau of Economic Research
The increasing ascendancy of real business
cycle theories of various stripes, with their
common view that the economy is best modeled
as a floating Walrasian equilibrium, buffeted by
productivity shocks, is indicative of the depths
of the divisions separating academic
macroeconomists. These theories deny
propositions thought self-evident by many
academic macroeconomists and all of those
involved in forecasting and controlling the
economy on a day-to-day basis. They assert that
monetary policies have no effect on real activity,
that fiscal policies influence the economy only
through their incentive effects, and that economic
fluctuations are caused entirely by supply rather
than demand shocks.
If these theories are correct, they imply that
the macroeconomics developed in the wake of
the Keynesian Revolution is well confined to
the ashbin of history. And they suggest that most
of the work of contemporary macroeconomists
is worth little more than that of those pursuing
astrological science. According to the views
espoused by enthusiastic proponents of real
business cycle theories, astrology and Keynesian
economics are in many ways similar: both lack
scientific support, both are premised on the
relevance of variables that are in fact irrelevant,
both are built on a superstructure of
nonoperational and ill-defined concepts, and both
are harmless only when they are ineffectual.
The appearance of Ed Prescott’s stimulating
paper, “Theory Ahead of Business Cycle
Measurement,” affords an opportunity to assess
the current state of real business cycle theory
and to consider its prospects as a foundation for
macroeconomic analysis. Prescott’s paper is
brilliant in highlighting the appeal of real business
cycle theories and making clear the assumptions
they require. But he does not make much effort
at caution in judging the potential of the real
business cycle paradigm. He writes that “if the
economy did not display the business cycle
phenomena, there would be a puzzle,”
characterizes without qualification economic
fluctuations as “optimal responses to uncertainty
in the rate of technological change,” and offers
the policy advice that “costly efforts at stabilization
are likely to be counter-productive.”
Prescott’s interpretation of his title is revealing
of his commitment to his theory. He does not
interpret the phrase theory ahead of measurement to
mean that we lack the data or measurements
necessary to test his theory. Rather, he means
that measurement techniques have not yet
progressed to the point where they fully
corroborate his theory. Thus, Prescott speaks of
the key deviation of observation from theory as
follows: “An important part of this deviation could
very well disappear if the economic variables
were measured more in conformity with theory.
That is why I argue that theory is now ahead of
business cycle measurement….”
The claims of real business cycle theorists
deserve serious assessment, especially given their
*An earlier version of these remarks was presented at the
July 25, 1986, meeting of the National Bureau of Economic
Research Economic Fluctuations Group.
23
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
98
FOUNDATIONS
source and their increasing influence within the
economics profession. Let me follow Prescott in
being blunt. My view is that real business cycle
models of the type urged on us by Prescott have
nothing to do with the business cycle phenomena
observed in the United States or other capitalist
economies. Nothing in Prescott’s papers or those
he references is convincing evidence to the contrary.
Before turning to the argument Prescott
presents, let me offer one lesson from the history
of science. Extremely bad theories can predict
remarkably well. Ptolemaic astronomy guided
ships and scheduled harvests for two centuries.
It provided extremely accurate predictions
regarding a host of celestial phenomena. And to
those who developed it, the idea that the earth
was at the center seemed an absolutely natural
starting place for a theory. So, too, Lamarckian
biology, with its emphasis on the inheritance of
acquired characteristics, successfully predicted
much of what was observed in studies of animals
and plants. Many theories can approximately
mimic any given set of facts; that one theory can
does not mean that it is even close to right.
Prescott’s argument takes the form of the
construction of an artificial economy which
mimics many of the properties of actual
economies. The close coincidence of his model
economy and the actual economy leads him to
conclude that the model economy is a reasonable
if abstract representation of the actual economy.
This claim is bolstered by the argument that the
model economy is not constructed to fit cyclical
facts but is parameterized on the basis of
microeconomic information and the economy’s
long-run properties. Prescott’s argument is
unpersuasive at four levels.
Are the Parameters Right?
First, Prescott’s claim to have parameterized the
model on the basis of well-established
microeconomic and long-run information is not
sustainable. As one example, consider a parameter
which Prescott identifies as being important in
determining the properties of the model, the share
of household time devoted to market activities.
He claims that is one-third. Data on its average
value over the last century indicate, as Martin
Eichenbaum, Lars Hansen, and Kenneth
Singleton (1986) have noted, an average value of
one-sixth over the past 30 years. This seems right—
a little more than half the adult population works,
and those who work work about a quarter of the
time. I am unable to find evidence supporting
Prescott’s one-third figure in the cited book by
Gilbert Ghez and Gary Becker (1975). To take
another example, Prescott takes the average real
interest rate to be 4 percent. Over the 30-year
period he studies, it in fact averaged only about 1
percent. This list of model parameters chosen
somewhat arbitrarily could be easily extended.
A more fundamental problem lies in Prescott’s
assumption about the intertemporal elasticity of
substitution in labor supply. He cites no direct
microeconomic evidence on this parameter,
which is central to his model of cyclical
fluctuations. Nor does he refer to any aggregate
evidence on it. Rather, he relies on a rather
selective reading of the evidence on the
intertemporal elasticity of substitution in
consumption in evaluating the labor supply
elasticity. My own reading is that essentially all
the available evidence suggests only a minimal
response of labor to transitory wage changes.
Many studies (including Altonji 1982; Mankiw,
Rotemberg, and Summers 1985; and
Eichenbaum, Hansen, and Singleton 1986)
suggest that the intertemporal substitution model
cannot account at either the micro or the macro
level for fluctuations in labor supply.
Prescott is fond of parameterizing models based
on long-run information. Japan has for 30 years
enjoyed real wage growth at a rate four times the
U.S. rate, close to 8 percent. His utility function
would predict that such rapid real wage growth
would lead to a much lower level of labor supply
by the representative consumer. I am not aware
that this pattern is observed in the data. Nor am I
aware of data suggesting that age/hours profiles
are steeper in professions like medicine or law,
where salaries rise rapidly with age.
Prescott’s growth model is not an
inconceivable representation of reality. But to
claim that its parameters are securely tied down
by growth and micro observations seems to me
a gross overstatement. The image of a big loose
tent flapping in the wind comes to mind.
Where Are the Shocks?
My second fundamental objection to Prescott’s
model is the absence of any independent
corroborating evidence for the existence of what
he calls technological shocks. This point is obviously
crucial since Prescott treats technological shocks
as the only driving force behind cyclical
fluctuations. Prescott interprets all movements
in measured total factor productivity as being
the result of technology shocks or to a small
extent measurement error. He provides no
discussion of the source or nature of these shocks,
nor does he cite any microeconomic evidence
for their importance. I suspect that the vast
majority of what Prescott labels technology
shocks are in fact the observable concomitants
of labor hoarding and other behavior which
Prescott does not allow in his model.
24
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
SCEPTICAL OBSERVATIONS
99
Lawrence H.Summers
Skeptical Observations
Two observations support this judgment.
First, it’s hard to find direct evidence of the
existence of large technological shocks. Consider
the oil shocks, certainly the most widely noted
and commented on shocks of the postwar period.
How much might they have been expected to
reduce total factor productivity? In one of the
most careful studies of this issue, Ernst Berndt
(1980, p. 85) concludes that “energy price or
quantity variations since 1973 do not appear to
have had a significant direct role in the slowdown
of aggregate labor productivity in U.S.
manufacturing, 1973–77.” This is not to deny
that energy shocks have important effects. But
they have not accounted for large movements in
measured total factor productivity.
Prescott assumes that technological changes
are irregular, but is unable to suggest any specific
technological shocks which presage the
downturns that have actually taken place. A
reasonable challenge to his model is to ask how
it accounts for the 1982 recession, the most serious
downturn of the postwar period. More generally,
it seems to me that the finding that measured
productivity frequently declines is difficult to
account for technologically. What are the sources
of technical regress? Between 1973 and 1977,
for example, both mining and construction
displayed negative rates of productivity growth.
For smaller sectors of the economy, negative
productivity growth is commonly observed.
A second observation casting doubt on
Prescott’s assumed driving force is that while
technological shocks leading to changes in total
factor productivity are hard to find, other
explanations are easy to support. Jon Fay and
James Medoff (1985) surveyed some 170 firms
on their response to downturns in the demand
for their output. The questions asked were
phrased to make clear that it was exogenous
downturns in their output that were being
inquired about. Fay and Medoff (1985, p. 653)
summarize their results by stating that “the
evidence indicates that a sizeable portion of the
swings in productivity over the business cycle
is, in fact, the result of firms’ decisions to hold
labor in excess of regular production
requirements and to hoard labor.” According to
their data, the typical plant in the U.S.
manufacturing sector paid for 8 percent more
blue-collar hours than were needed for regular
production work during the trough quarter of its
most recent downturn. After taking account of
the amount of other worthwhile work that was
completed by blue-collar employees during the
trough quarter, 4 percent of the blue-collar hours
paid for were hoarded. Similar conclusions have
been reached in every other examination of
microeconomic data on productivity that I am
aware of.
In Prescott’s model, the central driving force
behind cyclical fluctuations is technological shocks.
The propagation mechanism is intertemporal
substitution in employment. As I have argued so
far, there is no independent evidence from any
source for either of these phenomena.
What About Prices?…
My third fundamental objection to Prescott’s
argument is that he does price-free economic
analysis. Imagine an analyst confronting the
market for ketchup. Suppose she or he decided
to ignore data on the price of ketchup. This
would considerably increase the analyst’s
freedom in accounting for fluctuations in the
quantity of ketchup purchased. Indeed, without
looking at the price of ketchup, it would be
impossible to distinguish supply shocks from
demand shocks. It is difficult to believe that any
explanation of fluctuations in ketchup sales that
did not confront price data would be taken
seriously, at least by hard-headed economists.
Yet Prescott offers us an exercise in price-free
economics. While real wages, interest rates, and
returns to capital are central variables in his model,
he never looks at any data on them except for his
misconstrual of the average real interest rate over
the postwar period. Others have confronted
models like Prescott’s to data on prices with what
I think can fairly be labeled dismal results. There
is simply no evidence to support any of the price
effects predicted by the model. Prescott’s work
does not resolve—or even mention—the empirical
reality emphasized by Robert Barro and Robert
King (1982) that consumption and leisure move
in opposite directions over the business cycle with
no apparent procyclicality of real wages. It is
finessed by ignoring wage data. Prescott’s own
work with Rajnish Mehra (1985) indicates that
the asset pricing implications of models like the
one he considers here are decisively rejected by
nearly 100 years of historical experience. I simply
do not understand how an economic model can
be said to have been tested without price data.
I believe that the preceding arguments
demonstrate that real business cycle models of
the type surveyed by Prescott do not provide a
convincing account of cyclical fluctuations. Even
if this strong proposition is not accepted, they
suggest that there is room for factors other than
productivity shocks as causal elements in cyclical
fluctuations.
25
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
100
FOUNDATIONS
… And Exchange Failures?
A fourth fundamental objection to Prescott’s work
is that it ignores the fact that partial breakdowns
in the exchange mechanism are almost surely
dominant factors in cyclical fluctuations.
Consider two examples. Between 1929 and 1933,
the gross national product in the United States
declined 50 percent, as employment fell sharply.
In Europe today, employment has not risen since
1970 and unemployment has risen more than
fivefold in many countries. I submit that it defies
credulity to account for movements on this scale
by pointing to intertemporal substitution and
productivity shocks. All the more given that total
factor productivity has increased more than twice
as rapidly in Europe as in the United States.
If some other force is responsible for the
largest fluctuations that we observe, it seems
quixotic methodologically to assume that it plays
no role at all in other smaller fluctuations.
Whatever mechanisms may have had something
to do with the depression of the 1930s in the
United States or the depression today in Europe
presumably have at least some role in recent
American cyclical fluctuations.
What are those mechanisms? We do not yet
know. But it seems clear that a central aspect of
depressions, and probably economic fluctuations
more generally, is a breakdown of the exchange
mechanism. Read any account of life during the
Great Depression in the United States. Firms
had output they wanted to sell. Workers wanted
to exchange their labor for it. But the exchanges
did not take place. To say the situation was
constrained Pareto optimal given the
technological decline that took place between
1929 and 1933 is simply absurd, even though
total factor productivity did fall. What happened
was a failure of the exchange mechanism. This is
something that no model, no matter how
elaborate, of a long-lived Robinson Crusoe
dealing with his changing world is going to
confront. A model that embodies exchange is a
minimum prerequisite for a serious theory of
economic downturns.
The traditional Keynesian approach is to
postulate that the exchange mechanism fails
because prices are in some sense rigid, so they
do not attain market-clearing levels and thereby
frustrate exchange. This is far from being a
satisfactory story. Most plausible reasons why
prices might not change also imply that agents
should not continue to act along static demand
and supply curves. But it hardly follows that
ignoring exchange failures because we do not
yet fully understand them is a plausible strategy.
Where should one look for failures of the
exchange process? Convincing evidence of the
types of mechanisms that can lead to breakdowns
of the exchange mechanism comes from analyses
of breakdowns in credit markets. These seem to
have played a crucial role in each of the postwar
recessions. Indeed, while it is hard to account
for postwar business cycle history by pointing to
technological shocks, the account offered by, for
example, Otto Eckstein and Allen Sinai (1986)
of how each of the major recessions was caused
by a credit crunch in an effort to control inflation
seems compelling to me.
Conclusion
Even at this late date, economists are much better
at analyzing the optimal response of a single
economic agent to changing conditions than they
are at analyzing the equilibria that will result
when diverse agents interact. This unfortunate
truth helps to explain why macroeconomics has
found the task of controlling, predicting, or even
explaining economic fluctuations so difficult.
Improvement in the track record of
macroeconomics will require the development
of theories that can explain why exchange
sometimes works well and other times breaks
down. Nothing could be more counterproductive
in this regard than a lengthy professional detour
into the analysis of stochastic Robinson Crusoes.
26
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
SCEPTICAL OBSERVATIONS
101
Lawrence H.Summers
Skeptical Observations
References
Altonji, Joseph G. 1982. The intertemporal substitution model
of labour market fluctuations: An empirical analysis.
Review of Economic Studies 49 (Special Issue): 783–824.
Barro, Robert J., and King, Robert G. 1982. Time-separable
preferences and intertemporal-substitution models of
business cycles. Working Paper 888. National Bureau
of Economic Research.
Berndt, Ernst R. 1980. Energy price increases and the
productivity slowdown in United States
manufacturing. In The decline in productivity growth, pp.
60–89. Conference Series 22. Boston: Federal Reserve
Bank of Boston.
Eckstein, Otto, and Sinai, Allen. 1986. The mechanisms of
the business cycle in the postwar era. In The American
business cycle: Continuity and change, ed. Robert J.Gordon,
pp. 39–105. National Bureau of Economic Research
Studies in Business Cycles, vol. 25. Chicago: University
of Chicago Press.
Eichenbaum, Martin S.; Hansen, Lars P.; and Singleton,
Kenneth J. 1986. A time series analysis of representative
agent models of consumption and leisure choice under
uncertainty. Working Paper 1981. National Bureau
of Economic Research.
Fay, Jon A., and Medoff, James L. 1985. Labor and output
over the business cycle: Some direct evidence. American
Economic Review 75 (September): 638–55.
Ghez, Gilbert R., and Becker, Gary S. 1975. The allocation of
time and goods over the life cycle. New York: National Bureau
of Economic Research.
Mankiw, N.Gregory; Rotemberg, Julio J.; and Summers,
Lawrence H. 1985. Intertemporal substitution in
macroeconomics. Quarterly Journal of Economics 100
(February): 225–51.
Mehra, Rajnish, and Prescott, Edward C. 1985. The equity
premium: A puzzle. Journal of Monetary Economics 15
(March): 145–61.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
102
CHAPTER 6
Federal Reserve Bank of Minneapolis
Quarterly Review Fall 1986
Response to a Skeptic
Edward C.Prescott
Adviser
Research Department
Federal Reserve Bank of Minneapolis
and Professor of Economics
University of Minnesota
New findings in science are always subject to
skepticism and challenge. This is an important
part of the scientific process. Only if new results
successfully withstand the attacks do they
become part of accepted scientific wisdom.
Summers (in this issue) is within this tradition
when he attacks the finding I describe (in this
issue) that business cycles are precisely what
economic theory predicts given the best measures
of people’s willingness and ability to substitute
consumption and leisure, both between and
within time periods. I welcome this opportunity
to respond to Summers’ challenges to the
parameter values and the business cycle facts
that I and other real business cycle analysts have
used. In challenging the existing quality of
measurement and not providing measurement
inconsistent with existing theory, Summers has
conceded the point that theory is ahead of
business cycle measurement.
Miscellaneous Misfires
Before responding to Summers’ challenges to
the measurements used in real business cycle
analyses, I will respond briefly to his other attacks and, in the process, try to clarify some
methodological issues in business cycle theory
as well as in aggregate economic theory more
generally.
Prices
Summers asks, Where are the prices? This
question is puzzling. The mechanism real business cycle analysts use is the one he and other
leading people in the field of aggregate public
finance use: competitive equilibrium. Competitive equilibria have relative prices. As stated in
the introduction of “Theory Ahead of Business
Cycle Measurement” (in this issue), the business
cycle puzzle is, Why are there large movements
in the time allocated to market activities and
little associated movements in the real wage, the
price of people’s time? Along with that price,
Kydland and I (1982, 1984) examine the rental
price of capital. An infinity of other relative
prices can be studied, but these are the ones
needed to construct national income and product accounts. The behavior of these prices in
our models conforms with that observed.
In competitive theory, an economic
environment is needed. For that, real business
cycle analysts have used the neoclassical growth
model. It is the preeminent model in aggregate
economics. It was developed to account for the
growth facts and has been widely used for
predicting the aggregate effects of alternative tax
schemes as well. With the labor/leisure decision
endogenized, it is the appropriate model to
study the aggregate implications of technology
change uncertainty. Indeed, in 1977 Lucas, the
person responsible for making business cycles
again a central focus in economics, defined them
(p. 23) as deviations from the neoclassical
growth model—that is, fluctuations in hours
allocated to market activity that are too large to
be accounted for by changing marginal
productivities of labor as reflected in real wages.
Lucas, like me and virtually everyone else,
assumed that, once characterized, the
competitive equilibrium of the calibrated
neoclassical growth economy would display
much smaller fluctuations than do the actual
U.S. data. Exploiting advances in theory and
computational methods, Kydland and I (1982,
1984) and Hansen (1985) computed and
studied the competitive equilibrium process for
this model economy. We were surprised to find
28
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
RESPONSE TO A SCEPTIC
103
Edward C.Prescott
Response to a Skeptic
the predicted fluctuations roughly as large as
those experienced by the U.S. economy since
the Korean War.
Some economists have been reluctant to use
the competitive equilibrium mechanism to
study business cycle fluctuations because they
think it is contradicted by a real-world
observation: some individuals who are not
employed would gladly switch places with
similarly skilled individuals who are. Solow
(1986, p. S34), for example, predicted that “any
interesting and useful solution to that riddle will
almost certainly involve an equilibrium concept
broader, or at least different from, pricemediated market-clearing.” Rogerson (1984)
proved him wrong. If the world had no
nonconvexities or moral hazard problems,
Solow would be correct. But the mapping
between time allocated to market activities and
units of labor service produced does have
nonconvexities. Time spent commuting is not
producing labor services, yet it is time allocated
to market activity. With nonconvexities,
competitive equilibrium theory implies that the
commodities traded or priced are complicated
contracted arrangements which can include
employment lotteries with both winners and
losers. As shown by Hansen (1985),
competitive theory accounts well for the
observation that the principal margin of
adjustment in aggregate hours is the number of
people employed rather than the number of
hours worked per person—as well as for the
observation of so-called involuntary
unemployment.
Technology Shocks
Another Summers question is, Where are the
technology shocks? Apparently, he wants some
identifiable shock to account for each of the half
dozen postwar recessions. But our finding is not
that infrequent large shocks produce fluctuations; it is, rather, that small shocks do, every
period. At least since Slutzky (1927), some
stable low-order linear stochastic difference
equations have been known to generate cycles.
They do not have a few large shocks; they have
small shocks, one every period. The equilibrium allocation for the calibrated neoclassical
growth model with persistent shocks to technology turns out to be just such a process.
technology shocks as predicted by the neoclassical growth model. I do not argue that disruptions in the payment and credit system would
not disrupt the economy. That theory predicts
one factor has a particular nature and magnitude does not imply that theory predicts all
other factors are zero. I only claim that technology shocks account for more than half the fluctuations in the postwar period, with a best point
estimate near 75 percent. This does not imply
that public finance disturbances, random
changes in the terms of trade, and shocks to the
technology of exchange had no effect in that
period.
Neither do I claim that theory is ahead of
macroeconomic measurement in all respects.
As Summers points out, Mehra and I (1985)
have used the representative agent construct to
predict the magnitude of the average risk
premium of an equity claim over a real bill. Our
predicted quantity is small compared to the
historically observed average difference
between the yields of the stock market and U.S.
Treasury bills. But this is not a failure of the
representative agent construct; it is a success.
We used theory to predict the magnitude of the
average risk premium. That the representative
agent model is poorly designed to predict
differences in borrowing and lending rates—to
explain, for example, why the government can
borrow at a rate at least a few percentage points
less than the one at which most of us can
borrow—does not imply that this model is not
well designed for other purposes—for predicting
the consequences of technology shocks for
fluctuations at the business cycle frequencies,
for example.
Measurement Issues
Summers challenges the values real business
cycle analysts have selected for three model parameters. By arguing that historically the real
U.S. interest rate is closer to 1 percent than to
the model economy’s approximately 4 percent,
he is questioning the value selected for the subjective time discount factor. He explicitly questions our value for the leisure share parameter.
And Summers’ challenge to the observation
that labor productivity is procyclical is implicitly a challenge to my measure of the technology shock variance parameter.
My Claims
Real Interest Rate
Summers has perhaps misread some of my review of real business cycle research (in this issue). There I do not argue that the Great American Depression was the equilibrium response to
Summers points out that the real return on U.S.
Treasury bills over the last 30 years has been
about 1 percent, which is far from the average
real interest rate of the economies that Kydland
29
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
104
FOUNDATIONS
and I have studied. But for the neoclassical
growth model, the relevant return is not the return on T-bills. It is the return on tangible capital, such things as houses, factories, machines,
inventories, automobiles, and roads. The return
on capital in the U.S. business sector is easily
calculated from the U.S. National Income and
Product Accounts, so we use it as a proxy for the
return on U.S. capital more generally. This
number is obtained by dividing the income of
capital net of the adjusted capital consumption
allowance by the capital stock in the business
sector. For the postwar years, the result is approximately 4 percent, about the average real
return for the model economies.
Preferences
Summers also questions the value of the leisure
share parameter and argues that it is not well tied
down by micro observation at the household
level, as we claim. This is a potentially important parameter. If it is large, the response of labor
supply to temporary changes in the real wage is
large. Only if that response is large will large
movements in employment be associated with
small co-movements in the real wage.
Kydland and I conclude that the leisure
share parameter is not large based on findings
reported by Ghez and Becker (1975). They
report (p. 95) that the annual productive time
endowment of U.S. males is 5,096 hours. They
also say (p. 95) that U.S. females allocate about
75 hours per week to personal care, leaving 93
hours of production time per week. This
multiplied by 52 is 4,836 hours, the annual
productive time endowment of females. Ghez
and Becker also report the average annual hours
of employment for noninstitutionalized,
working-age males as about 2,000 hours (pp.
85–91). If females allocate half as many hours to
market employment as do males, the average
fraction of time the U.S. working-age
population spends in employment is about
0.30. Adding to this the time spent commuting
yields a number close to those for our models.
(They are all between 0.30 and 0.31 in Kydland
and Prescott 1982 and 1984.)
Initially Kydland and I used time additive
preferences, and the predictions of theory for
productivity movements were as large in
percentage terms as aggregate hour movements.
This is inconsistent with observations, so I did
not take seriously the prediction of theory that a
little over half the aggregate output fluctuations
in the postwar period were responses to
technology shocks. At that time, measurement
was still ahead of theory. Then, the prediction of
theory would have been consistent with the
relative movement of productivity and aggregate
hours, and technology shocks would have
accounted for the business cycle phenomena, if
the leisure share parameter were five-sixths. With
the discipline we used, however, this share
parameter had to be consistent with observations
on household time allocation. That we are now
debating about a theory of aggregate phenomena
by focusing on household time allocation is
evidence that economic theory has advanced.
Now, like physical scientists, when economists
model aggregate phenomena, the parameters
used can be measured independently of those
phenomena.
In our 1982 paper, Kydland and I did claim
that fluctuations of the magnitude observed
could plausibly be accounted for by the
randomness in the technological change
process. There we explored the implications of
a distributed lag of leisure being an argument of
the period utility function rather than just the
current level of leisure. Like increasing the
leisure share parameter, this broadening results
in larger fluctuations in hours in response to
technology shocks. Kydland (1983) then
showed that an unmeasured household-specific
capital stock could rationalize this distributed
lag. In addition, the lag was not inconsistent
with good micro measurement, and these
parameters could be measured independently
of the business cycle phenomena. The
distributed lag was a long shot, though, so we
did not claim that theory had caught up to
measurement.
Since then, however, two panel studies
found evidence for a distributed lag of the type
we considered (Hotz, Kydland, and Sedlacek
1985; Eckstein and Wolpin 1986). With this
development, theory and measurement of the
business cycle became roughly equal.
Subsequently, an important advance in
aggregate theory has made moot the issue of
whether Kydland’s and my assumed
preferences for leisure are supported by micro
measurement. Given an important
nonconvexity in the mapping between time
allocated to market activities and units of labor
service produced, Rogerson (1984) showed
that the aggregate elasticity of labor supply to
temporary changes in the real wage is large
independent of individuals’ willingness to
intertemporally substitute leisure. This nicely
rationalized the disparate micro and macro
labor findings for this elasticity—the
microeconomists’ that it is small (for example,
Ashenfelter 1984) and the macroeconomists’
that it is large (for example, Eichenbaum,
Hansen, and Singleton 1984). Hansen (1985)
30
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
RESPONSE TO A SCEPTIC
105
Edward C.Prescott
Response to a Skeptic
introduced this nonconvexity into the
neoclassical growth model. He found that with
this feature theory predicts that the economy
will display the business cycle phenomena even
if individuals’ elasticity of labor supply to
temporary changes in the real wage is small.
Further, with this feature he found theory
correctly predicts that most of the variation in
aggregate hours of employment is accounted for
by variation in the number of people employed
rather than in the number of hours worked per
person.
Technology
… Uncertainty
In our 1982 paper, Kydland and I searched
over processes for the technological change
process. We did sensitivity analysis with the
other parameters, but found the conclusions
relatively insensitive to their assumed values
(except for the distributed lag of leisure parameters just discussed). The parameters of the technological change process did affect our predictions of the aggregate implications of uncertainty in the technology parameter. In fact,
Lucas (1985, p. 48) criticized us for searching
for the best fit. In “Theory Ahead of Business
Cycle Measurement,” I directly examined the
statistical properties of the technology coefficient process. I found that the process is an approximate random walk with standard deviation of change in the logs approximately
0.00763 per quarter. When this number is used
in the Hansen model, fluctuations predicted are
even larger than those observed. In Kydland’s
and my model (1984), they are essentially equal
to those observed.
Some, on the basis of theory, think that the
factors producing technological change are
small, many, and roughly uncorrelated. If so,
by the law of large numbers, these factors
should average out and the technological
change process should be very smooth. I found
(in this issue) empirical evidence to the
contrary. Others have too. Summers and
Heston (1984) report the annual gross national
products for virtually every country in the
postwar period. They show huge variation
across countries in the rate of growth of per
capita income over periods sufficiently long
that business cycle variations are a minor
consideration. Even individual countries have
large variation in the decade growth rates of per
capita output. Given Solow’s (1957) finding
that more than 75 percent of the changes in per
capita output are accounted for by changes in
the technology parameter, the evidence for
variation in the rate of technological advance is
strong.
Obviously, economists do not have a good
theory of the determinants of technological
change. In this regard, measurement is ahead of
theory. The determinants of the rate of
technological change must depend greatly on
the institutions and arrangements that societies
adopt. Why else should technology advance
more rapidly in one country than in another or,
within a country, more rapidly in one period
than in another? But a theory of technological
change is not needed to predict responses to
technological change.
The key parameter is the variance of the
technology shock. This is where better
measurement could alter the prediction of
theory. Is measuring this variance with Solow’s
(1957) method (as I did) reasonable? I showed
that measures of the technology shock variance
are insensitive to cyclical variations in the
capital utilization rate. Even if that rate varies
proportionately to hours of employment and
the proportionality constant is selected so as to
minimize the measured standard deviation of
the technology shock, that measured deviation
is reduced only from 0.00763 to 0.00759.
Further, when the capital utilization rate varies
in this way for the model, the equilibrium
responses are significantly larger. Variation in
the capital utilization rate does not appear to
greatly bias my estimate of the importance of
technological change variance for aggregate
fluctuations.
Perhaps better measurement will find that the
technological change process varies less than I
estimated. If so, a prediction of theory is that the
amount of fluctuation accounted for by
uncertainty in that process is smaller. If this were
to happen, I would be surprised. I can think of
no plausible source of measurement error that
would produce a random walk-like process for
technological change.
… Labor Hoarding
Summers seems to argue that measured
productivity is procyclical because measurement
errors are cyclical. To support his argument, he
cites a survey by Fay and Medoff (1985), which
actually has little if anything to say about cyclical
movements. Fay and Medoff surveyed more than
1,000 plant managers and received 168 usable
responses. One of the questions asked was, How
many extra blue-collar workers did you have in
your recent downturn? They did not ask, How
many extra workers did you have at the trough
quarter and at the peak quarter of the most recent
business cycle? Answers to those questions are
30
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
106
FOUNDATIONS
needed to conclude how the number of extra
blue-collar workers reported by managers varies
over the cycle. Even if these questions had been
asked, though, the response to them would not
be a good measure of the number of redundant
workers. Such questions are simply too
ambiguous for most respondents to interpret
them the same way.
The argument that labor hoarding is cyclical
is not supported by theory either. The fact that
labor is a quasi-fixed factor of production in the
sense of Oi (1962) does not imply that more
workers will be hoarded in recessions than in
expansions. In bad times a firm with low output
may be less reluctant to lay off workers than in
good times because the worker is less likely to
be hired by another firm. This argument
suggests that labor hoarding associated with
firm-specific output variations should be
procyclical. Leisure consumed on the job also
may be less in bad times than in good because
work discipline may be greater. That is, an
entrepreneur might be less reluctant to fire a
worker in bad times because the worker can
more easily be replaced. One might reasonably
think, therefore, that labor’s quasi-fixed nature
makes measured productivity less, not more,
cyclically volatile than productivity really is.
There is another, better reason to think that.
In the standard measures of aggregate hours of
employment, the hours of an experienced MBA
from one of the best business schools are treated
the same as those of a high school dropout. Yet
these hours do not on average command the
same price in the market, which is evidence that
they are not the same commodity. In the
neoclassical growth model, the appropriate way
to aggregate hours is in terms of effective units of
labor. That is, if the MBA’s productivity is five
times that of the high school dropout, then each
hour of the MBA’s time is effectively equivalent
to five hours of the high school dropout’s time.
The work of Kydland (1984) suggests this
correction is an important one. The more
educated and on average more highly paid have
much less variability in annual hours of
employment than do the less educated.
Kydland (1984, p. 179) reports average hours
and average wages as well as sensitivity of hours
to the aggregate unemployment rate for adult
males categorized by years of schooling. His
figures imply that a 1 percentage point change
in the aggregate unemployment rate for adult
males is associated with a 1.24 percent change
in equally weighted hours. When those hours
are measured as effective units of labor, the
latter change is only 0.65 percent. This is strong
evidence that if the labor input were measured
correctly, the measure of productivity would
vary more.
To summarize, measurement of the labor
input needs to be improved. By questioning the
standard measures, Summers is agreeing that
theory is ahead of business cycle measurement.
More quantitative theoretic work is also
needed, to determine whether abstracting from
the fact that labor is a partially fixed factor
affects any of the real business cycle models’
findings. Of course, introducing this feature—or
others—into these models may significantly alter
their predictions of the aggregate implications of
technology uncertainty. But respectable
economic intuition must be based on models
that have been rigorously analyzed.
To Conclude
Summers cannot be attacking the use of competitive theory and the neoclassical growth environment in general. He uses this standard
model to predict the effects of alternative tax
policies on aggregate economic behavior. He
does not provide criteria for deciding when implications of this model should be taken seriously and when they should not be. My guess is
that the reason for skepticism is not the methods
used, but rather the unexpected nature of the
findings. We agree that labor input is not that
precisely measured, so neither is technological
uncertainty. In other words, we agree that
theory is ahead of business cycle measurement.
31
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
RESPONSE TO A SCEPTIC
107
Edward C.Prescott
Response to a Skeptic
References
Ashenfelter, O. 1984. Macroeconomic analyses
and microeconomic analyses of labor
supply. In Essays on macroeconomic Implications
of financial and labor markets and political
processes, ed. K.Brunner and A.H.Meltzer.
Carnegie-Rochester Conference Series on
Public Policy 21:117–55. Amsterdam:
North-Holland.
Eckstein, Zvi, and Wolpin, Kenneth I. 1986.
Dynamic labor force participation of
married women and endogenous work
experience. Manuscript. Tel Aviv
University and Ohio State University.
Eichenbaum, Martin S.; Hansen, Lars P.; and
Singleton, Kenneth J. 1984. A time series
analysis of representative agent models of
consumption and leisure choice under
uncertainty. Working paper, CarnegieMellon University.
Fay, Jon A., and Medoff, James L. 1985. Labor
and output over the business cycle: Some
direct evidence. American Economic Review
75 (September): 638–55.
Ghez, Gilbert R., and Becker, Gary S. 1975. The
allocation of time and goods over the life cycle.
New York: National Bureau of Economic
Research.
Hansen, Gary D. 1985. Indivisible labor and
the business cycle. Journal of Monetary
Economics 16 (November): 309–27.
Hotz, V.S.; Kydland, F.E.; and Sedlacek, G.L.
1985. Intertemporal preferences and labor
supply. Working paper, Carnegie-Mellon
University.
Kydland, Finn E. 1983. Nonseparable utility and
labor supply. Working paper, Hoover
Institution.
____. 1984. Labor-force heterogeneity and the
business cycle. In Essays on macroeconomic
implications of financial and labor markets and
political processes, ed. K.Brunner and
A.H.Meltzer.
Carnegie-Rochester
Conference Series on Public Policy 21:173–
208. Amsterdam: North-Holland.
Kydland, Finn E., and Prescott, Edward C. 1982.
Time to build and aggregate fluctuations.
Econometrica 50 (January): 1345–70.
____. 1984. The workweek of capital and labor.
Research Department Working Paper 267.
Federal Reserve Bank of Minneapolis.
Lucas, Robert E., Jr. 1977. Understanding
business cycles. In Stabilization of the domestic
and international economy, ed. K.Brunner and
A.H.Meltzer.
Carnegie-Rochester
Conference Series on Public Policy 5:7–
29. Amsterdam: North-Holland.
____. 1985. Models of business cycles.
Manuscript prepared for the Yrjo Jahnsson
Lectures , Helsinki, Finland. University of
Chicago.
Mehra, Rajnish, and Prescott, Edward C. 1985.
The equity premium: A puzzle. Journal of
Monetary Economics 15 (March): 145–61.
Oi, Walter Y. 1962. Labor as a quasi-fixed factor.
Journal of Political Economy 70 (December):
538–55.
Rogerson, R.D. 1984. Indivisible labor, lotteries
and equilibrium. Economics Department
Working Paper 10. University of Rochester.
Slutzky, Eugen. 1927. The summation of random
causes as the source of cyclic processes. In
Problems of economic conditions, ed.
Conjuncture Institute, Moskva (Moscow),
vol. 3, no. 1. Revised English version, 1937,
in Econometrica 5:105–46.
Solow, Robert M. 1957. Technical change and
the aggregate production function. Review
of Economics and Statistics 39 (August): 312–
20.
_____. 1986. Unemployment: Getting the
questions right. Economica 53 (Supplement):
S23–S34.
Summers, Robert, and Heston, Alan. 1984.
Improved international comparisons of real
product and its composition: 1950–1980.
Review of Income and Wealth 30 (June): 207–62.
33
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
108
CHAPTER 7
Journal of Monetary Economics 21 (1988) 195–232. North-Holland
PRODUCTION, GROWTH AND BUSINESS CYCLES
I. The Basic Neoclassical Model
Robert G.KING, Charles I.PLOSSER and Sergio T.REBELO*
University of Rochester, Rochester, NY 14627, USA
Received September 1987, final version received December 1987
This paper presents the neoclassical model of capital accumulation augmented by choice
of labor supply as the basic framework of modern real business cycle analysis. Preferences
and production possibilities are restricted so that the economy displays steady state
growth. Then we explore the implications of the basic model for perfect foresight capital
accumulation and for economic fluctuations initiated by impulses to technology. We
argue that the neoclassical approach holds considerable promise for enhancing our
understanding of fluctuations. Nevertheless, the basic model does have some important
shortcomings. In particular, substantial persistence in technology shocks is required
if the model economy is to exhibit periods of economic activity that persistently deviate
from a deterministic trend.
1. Introduction and summary
Real business cycle analysis investigates the role of neoclassical factors in
shaping the character of economic fluctuations. In this pair of essays, we
provide an introduction to the real business cycle research program by
considering the basic concepts, analytical methods and open questions
on the frontier of research. The focus of the present essay is on the dynamic
aspects of the basic neoclassical model of capital accumulation. This model
is most frequently encountered in analyses of economic growth, but we
share Hicks’ (1965, p. 4) perspective that it is also a basic laboratory for
investigating more general dynamic phenomena involving the choice of
consumption, work effort and investment.
Our use of the neoclassical model of capital accumulation as the engine
of analysis for the investigation of economic fluctuations raises a number of
central issues. First, what role does economic growth play in the study of
* The authors acknowledge financial support from the National Science Foundation.
King and Plosser have joint affiliations with the Department of Economics and the
W.E.Simon Graduate School of Business, University of Rochester. Rebelo is affiliated
with the Department of Economics, University of Rochester and the Department of
Economics, Portuguese Catholic University. We have benefited from the comments
of Andrew Abel and Larry Christiano, as well as from those of seminar participants
at the Federal Reserve Bank of Richmond, Brasenose College, Oxford, Institute for
International Economic Studies, University of Stockholm, Northwestern University,
Yale University, and Columbia University.
0304–3932/88/$3.50 ©1988, Elsevier Science Publishers B.V. (North-Holland)
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE BASIC NEOCLASSIC MODEL
196
109
R.G.King et al., Production, growth and business cycles I
economic fluctuations? More precisely, does the presence of economic growth
restrict the preference and production specifications in ways that are important
for the analysis of business cycles? Second, what analytical methods can be
employed to study the time series implications of the neoclassical model?
Third, what are the dynamics of the neoclassical model in response to
technology shocks? Finally, does the neoclassical model—driven by
technology shocks—replicate important features of macroeconomic time
series? The analysis of these issues forms the core of the present paper and
establishes the building blocks of real business cycle theory.
Real business cycle theory, though still in the early stages of
development, holds considerable promise for enhancing our understanding
of economic fluctuations and growth as well as their interaction. The
basic framework developed in this essay is capable of addressing a wide
variety of issues that are commonly thought to be important for
understanding business cycles. While we focus here on models whose
impulses are technological, the methods can be adapted to consider shocks
originating from preferences or other exogenous factors such as government
policies and terms of trade. Some of these extensions to the basic framework
are developed in the companion essay.
To many readers it must seem heretical to discuss business cycles without
mentioning money. Our view, however, is simply that the role of money
in an equilibrium theory of economic growth and fluctuations remains
an open area for research. Further, real disturbances generate rich and
neglected interactions in the basic neoclassical model that may account
for a substantial portion of observed fluctuations. The objective of real
business cycle research is to obtain a better understanding of the character
of these real fluctuations. Without an understanding of these real
fluctuations it is difficult a priori to assign an important role to money.
The organization of the paper follows the sequence of questions outlined
above. We begin in section 2 by describing the preferences, endowments and
technology of the basic (one-sector) neoclassical model of capital accumulation.1
In contrast to the familiar textbook presentation of this model, however, work
effort is viewed as a choice variable. We then discuss the restrictions on
production possibilities and preferences that are necessary for steady state
growth. On the production side, with a constant returns to scale production
function, technical progress must be expressible in labor augmenting (Harrod
neutral) form. In a feasible steady state, it follows that consumption, investment,
output and capital all must grow at the exogenously specified rate of technical
change. On the other hand, since the endowment of time is constant, work
effort cannot grow in the steady state. Thus, preferences must be restricted so
that there is no change in the level of effort on the steady state growth path
despite the rise in marginal productivity stemming from technical progress,
1
A more detailed and unified development of the material is presented in the
technical appendix, available from the authors on request.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
110
FOUNDATIONS
R.G.King et al., Production, growth and business cycles I
197
i.e., there must be an exact offset of appropriately defined income and
substitution effects.
Section 3 concerns perfect foresight dynamic competitive equilibria,
which we analyze using approximations near the steady state. Using a
parametric version of the model, with parameters chosen to match the
long-run U.S. growth experience, we study the interaction between
intertemporal production possibilities and the equilibrium quantity of
labor effort. Off the steady state path, we find that capital and effort are
negatively related despite the fact that the marginal product of labor
schedule is positively related to the capital stock. That is, in response to
the high real rate of return implied by a low capital stock, individuals will
substitute intertemporally to produce additional resources for investment.
Working from a certainty equivalence perspective, section 4 considers
how temporary productivity shocks influence economic activity, generating
‘real business cycles’ in the terminology of Long and Plosser (1983). Again
there is an important interaction between variation in labor input—this
time in response to a productivity shock—and the intertemporal substitution
in production permitted by capital accumulation. Purely temporary
technology shocks call forth an expansion of labor input once the Long
and Plosser (1983) assumption of complete depreciation is replaced by a
more realistic value,2 since more durable capital increases the feasibility of
intertemporal substitution of goods and leisure. Nevertheless, with purely
temporary productivity shocks, we find that there are important deficiencies
of the basic neoclassical model. Although there is substantial serial
correlation in consumption and capital as a consequence of consumption
smoothing, there is effectively no serial correlation in output or employment.
This lack of propagation reflects two basic properties of the parameterized
model: (i) a negative relation between capital and effort along the transition
path and (ii) the minor effect of a purely temporary technology shock on
a large and durable capital stock. Thus, the basic neoclassical capital
accumulation mechanism is important for permitting intertemporal
substitution of goods and leisure, but it does not generate serial correlation
in output and employment close to that exhibited by macroeconomic data.
It is necessary, therefore, to incorporate substantial serial correlation in
productivity shocks [as in Kydland and Prescott (1982), Long and Plosser
(1983), Hansen (1985), and Prescott (1986)] if the basic neoclassical model
is to generate business fluctuations that resemble those in post-war U.S.
experience. Since serial correlation involves movements in productive
opportunities that are more persistent in character, labor input responds
less elastically to a given size shock, but its response remains positive. On
the other hand, with more persistent productivity shocks, consumption
responds more elastically in accord with the permanent income theory.
2
By a purely temporary shock, we mean one that lasts for a single time period,
which is taken to be a quarter in our analysis.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE BASIC NEOCLASSIC MODEL
198
111
R.G.King et al., Production, growth and business cycles I
In section 5, we show that the basic neoclassical model—with persistent
technology shocks—captures some key features of U.S. business cycles.
For example, the model replicates observed differences in volatility across
key series. Measured as a percentage of the standard deviation of output,
there is an identical ordering of the model’s implications for investment,
wages, consumption and hours, and the U.S. time series: investment is
most volatile, followed by wages, consumption and then hours. But there
are also aspects of the data that are poorly captured by the single-shock
model. For example, consumption, investment and hours are much more
highly correlated with output in the model than in the data.
Professional interest in real business cycle analysis has been enhanced
by the comparison of moments implied by neoclassical models with those
of U.S. time series, as initiated by Kydland and Prescott (1982). Our
implications for moments differ from those of Hansen (1985) and Prescott
(1986), principally because we do not filter actual and model-generated
time series to remove slow-moving components. For example, in Hansen’s
and Prescott’s analyses, filtered hours and output have virtually identical
volatilities, in both the model and the transformed data. By contrast, in
our analysis, the volatility of hours is about half that of output (both in
our model and post-war detrended U.S. data). These differences occur
despite the fact that there is little economic difference in the models under
study.
Section 6 provides a brief summary and some concluding remarks.
2. The basic neoclassical model
Our analysis of economic growth and fluctuations starts by summarizing
the key features of the basic one-sector, neoclassical model of capital
accumulation. Much of the discussion in this section will be familiar to
readers of Solow (1956), Cass (1965), Koopmans (1965) and subsequent
textbook presentations of their work, but it is important to build a base for
subsequent developments.
2.1. Economic environment
We begin by considering the preferences, technology and endowments of
the environment under study.
Preferences. We consider an economy populated by many identical infinitelylived individuals with preferences over goods and leisure represented by
(2.1)
where Ct is commodity consumption in period t and Lt is leisure in period
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
112
FOUNDATIONS
R.G.King et al., Production, growth and business cycles I
199
t. Consumption and leisure are assumed throughout to be goods, so that
utility is increasing in Ct and Lt.3
Production possibilities. There is only one final good in this economy and it
is produced according to a constant returns to scale neoclassical production
technology given by
(2.2)
where Kt is the predetermined capital stock (chosen at t-1) and Nt is the
labor input in period t.4 We permit temporary changes in total factor
productivity through At. Permanent technological variations are restricted
to be in labor productivity, Xt, for reasons that we discuss below.
Capital accumulation. In this simple neoclassical framework the commodity
can be either consumed or invested. The capital stock evolves according to
(2.3)
where It is gross investment and δK is the rate of depreciation of capital.5
Resource constraints. In each period, an individual faces two resource
constraints: (i) total time allocated to work and leisure must not exceed
the endowment, which is normalized to one, and (ii) total uses of the
commodity must not exceed output. These conditions are
(2.4)
(2.5)
Naturally, there are also the non-negativity constraints Lt≥0, Nt≥0, Ct≥0
and Kt≥0.
3
Momentary utility, u(·), is assumed to be strictly concave and twice continuously
differentiable. Further, it satisfies the Inada conditions, namely that limc→0D1u(c,
L)=∞ and limc→∞D1u(c, L)=0, limL→0D2u(c, L)=∞ and limL→1D2u(c, L)=0, where Diu(·)
is the first partial derivative of u(·) with respect to the function’s i th argument.
4
By neoclassical, we mean that the production function is concave, twice
continuously differentiable, satisfies the Inada conditions, and that both factors are
essential in production.
5
We abstract from adjustment costs to capital accumulation throughout the analysis,
as these seem to us to be basically a restricted formulation of the two-sector neoclassical
model.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE BASIC NEOCLASSIC MODEL
200
113
R.G.King et al., Production, growth and business cycles I
2.2. Individual optimization and competitive equilibrium
The standard neoclassical analysis focuses on the optimal quantities chosen by
a ‘social planner’ or representative agent directly operating the technology of
the economy. Since our setup satisfies the conditions under which the second
welfare theorem is valid, optimal capital accumulation will also be realized in a
competitive equilibrium.6 In the companion essay, we discuss departures from
the strict representative agent model including government expenditures and
distorting taxes, productive externalities, and heterogeneity of preferences and
productivities. In these contexts, we will need to be more precise about
distinguishing between individual choices and competitive outcomes.
2.3. Steady state growth
A characteristic of most industrialized economies is that variables like output
per capita and consumption per capita exhibit sustained growth over long
periods of time. This long-run growth occurs at rates that are roughly
constant over time within economies but differ across economies. We interpret
this pattern as evidence of steady state growth, by which we mean that the
levels of certain key variables grow at constant—but possibly different—rates,
at least some of which are positive. Additional restrictions on preferences
and technologies are required if the system is to exhibit steady state growth.
Restrictions on production. For a steady state to be feasible, Swan (1963) and
Phelps (1966) show that permanent technical change must be expressible
in a labor augmenting form, which rationalizes our specification in (2.2)
above. To make for an easier comparison with other studies, we adopt the
Cobb-Douglas production process for the bulk of our analysis,
(2.6)
where the quantity NtXt is usually referred to as effective labor units.7
Since variation in At is assumed temporary, we can ignore it for our
investigation of steady state growth. The production function (2.6) and
the accumulation equation (2.3) then imply that the steady state rates of
growth of output, consumption, capital and investment per capita are all
equal to the growth rate of labor augmenting technical progress.8 Denoting
6
The basic reference is Debreu (1954). See also Prescott and Lucas (1972).
We note, however, that if technological change is labor augmenting, then the
observed invariance of factor shares to the scale of economic activity cannot be used
to rationalize the restriction to the Cobb-Douglas form. In the presence of labor
augmenting technological progress the factor shares are constant for any constant
returns to scale production function.
7
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
114
FOUNDATIONS
R.G.King et al., Production, growth and business cycles I
201
one plus the growth rate of a variable Z as γz (i.e., Zt+1/Zt), then any
feasible steady state requires
(2.7a)
and the growth rate of work effort to be zero, i.e.,
(2.7b)
Since time devoted to work N is bounded by the endowment, it cannot
grow in the steady state (2.7b). Thus, the only admissible constant growth
rate for N is zero.
In any such feasible steady state, the marginal product of capital and
the marginal product of a unit of labor input in efficiency units are constant.
The levels of the marginal products, however, depend on the ratio of
capital to effective labor, which is not determined by the restriction to a
feasible steady state.
Restrictions on preferences. Eqs. (2.7a) and (2.7b) describe the technologically
feasible steady state growth rates. If these conditions are not compatible
with the efficiency conditions of agents in the economy, then they are of
little interest since they would never be an equilibrium outcome. We can
insure that the feasible steady state is compatible with an (optimal)
competitive equilibrium, however, by imposing two restrictions on
preferences: (i) the intertemporal elasticity of substitution in consumption
must be invariant to the scale of consumption and (ii) the income and
substitution effects associated with sustained growth in labor productivity
must not alter labor supply.
The first condition must hold because the marginal product of capital,
which equals one plus the real interest rate in equilibrium, must be constant
in the steady state. Since consumption is growing at a constant rate and
the ratio of discounted marginal utilities must equal one plus the interest
rate, the intertemporal elasticity of substitution must be constant and
independent of the level of consumption.
The second condition is required because hours worked cannot grow
(γN=1) in the steady state. To reconcile this with a growing marginal
productivity of labor—induced by labor augmenting technical change (Xt)—
income and substitution effects of productivity growth must have exactly
offsetting effects on labor supply (N).9
8
9
This result, in fact, holds for any constant returns to scale production function.
Effective labor (NXt) will continue to grow at rate γx.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE BASIC NEOCLASSIC MODEL
8
202
115
R.G.King et al., Production, growth and business cycles I
These conditions imply the following class of admissible utility
functions:10
(2.8a)
for 0<<1 and >1, while for =1,
(2.8b)
Some additional restrictions are necessary to assure that (i) consumption
and leisure are goods and (ii) that utility is concave.11 The constant
intertemporal elasticity of substitution in consumption is 1/ for these
utility functions. For the remainder of our analysis we restrict ourselves
to utility functions of this class.
The requirement that preferences be compatible with steady state growth
has important implications for the study of economic fluctuations. If there
is no capital [i.e., if the production function is just of the form At(NtXt)],
then there will be no response of hours to variation in Xt or At in general
equilibrium. This arises because (i) utility implies that the income and
substitution effects of wage changes just offset and (ii) with no intertemporal
substitution in production, income effects must be fully absorbed within
any decision period [as in Barro and King (1984)]. Thus, in all of the
parameterizations of the neoclassical model that we consider, variations
in work effort are associated with intertemporal substitution made possible
in equilibrium by capital accumulation.
2.4. Stationary economies and steady states
The standard method of analyzing models with steady state growth is to
transform the economy into a stationary one where the dynamics are
more amenable to analysis. In the context of the basic neoclassical model,
this transformation involves dividing all variables in the system by the
growth component X, so that c=C/X, k=K/X, i=I/X, etc. This economy is
identical to a simple ‘no-growth’ economy with two exceptions. First the
capital accumulation equation, Kt+1=(1–δK)Kt+It, becomes γXkt+1=(1-δK)kt+it.
Second, transforming consumption in the preference specification generally
alters the effective rate of time preference. That is,
10
See the technical appendix for a demonstration of the necessity of these conditions
and that they imply (2.8a) and (2.8b).
11
When momentary utility is additively separable (2.8b), all that we require is that (L)
is increasing and concave. When momentary utility is multiplicatively separable, then we
require that (L) is (i) increasing and concave if <1 and (ii) decreasing and convex if >1.
Defining Dn(L) as the n th total derivative of the function (L), we further require that—
[LD2(L)/D(L)]>(1-)[LD(L)/(L)] to assure overall concavity of u(·).
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
116
FOUNDATIONS
R.G.King et al., Production, growth and business cycles I
203
(2.9a)
(2.9b)
where β*=β(γx)1-σ and β*<1 is required throughout to guarantee finiteness
of lifetime utility. Thus, unless =1, β*≠β. By suitable selection of X0, we
Combining the
can in either case make the objective
resource constraints, we form the Lagrangian
(2.10)
The efficiency conditions for the transformed economy are (2.11)–
(2.15). In these expressions, Di is the first partial derivative operator with
respect to the ith argument. For convenience, we discount the Lagrange
multipliers, i.e.
(2.11)
(2.12)
(2.13)
(2.14)
(2.15)
where (2.11)–(2.14) must hold for all t=1, 2, …∞ and (2.15) is the so-called
transversality condition. The economy’s initial capital stock, k0, is given.
Optimal per capita quantities for this economy—for a given sequence
of technology shifts—are sequences of consumption
effort
capital stock
and shadow prices
that satisfy the
efficiency conditions (2.11)–(2.15). Under our assumptions about
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE BASIC NEOCLASSIC MODEL
204
117
R.G.King et al., Production, growth and business cycles I
preferences and production possibilities, conditions (2.11)–(2.15) are
necessary and sufficient for an optimum.12
The prices that decentralize the optimal solution as a competitive equilibrium
and the optimal sequences
can be computed using the technology shifts
and
For instance, in a complete initial date markets
framework the sequence of equilibrium prices of labor and the final good are,
and
Under perfect foresight
respectively,
(rational expectations), a regime of sequential loan markets and spot markets
in labor services also supports the optimal solution as a competitive equilibrium.
In this market structure, the relevant prices are the real interest rate between
t and t+1, rt, and the real wage rate, wt. It is easy to demonstrate that these are
given by (1+rt)=γXλt/λt+1β* and wt=AtD2F(kt, Nt).
3. Perfect foresight capital accumulation
A major feature of the basic one sector neoclassical model with stationary
technology is that the optimal capital stock converges monotonically to a
stationary point.13 While qualitative results such as the monotonicity
property are important, we wish to undertake quantitative analyses of
capital stock dynamics. This requires that we exploit the fact that (2.11)–
(2.14) can be reduced to a non-linear system of first-order difference
equations in k and λ or a second-order equation in k only. The two
boundary conditions of this system are the transversality condition (2.15)
and the initial capital stock, k0. We focus on approximate linear dynamics
in the neighborhood of the steady state denoted by (A, k, N, c and y).14
3.1. Approximation method
The initial step in obtaining the system of linear difference equations is to
approximate (2.11)–(2.14) near the stationary point. To do this, we express
each condition in terms of the percentage deviation from the
stationary value, which we indicate using a circumflex
Then, we linearize each condition
12
See Weitzman (1973) and Romer and Shinotsuka (1987).
In the fixed labor case, which is the most thoroughly studied, this property has
been shown to derive principally from preferences, in that the concavity of u(·) is
sufficient for monotonicity so long as there is a maximum sustainable capital stock
[Boyd (1986) and Becker et al. (1986)]. In environments such as ours, where the
production function is strictly concave in capital (for fixed labor), monotonicity also
insures that capital approaches a unique stationary point.
14
The technical appendix discusses solution methods in considerable detail. The
linear approximation method, it should be noted, rules out certain phenomena that
may arise in the basic neoclassical model, such as a humped shaped transition path for
investment [see King (1987)].
13
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
118
FOUNDATIONS
R.G.King et al., Production, growth and business cycles I
205
in terms of deviations from the stationary point. The results for the first
two conditions can be written as follows:
(3.1)
(3.2)
where ξab is the elasticity of the marginal utility of a with respect to b.15
Approximation of the intertemporal efficiency condition (2.13) implies that
(3.3)
where ηA is the elasticity of the gross marginal product of capital with
respect to A evaluated at the steady state, etc.16 Finally, approximation of
the resource constraint (2.14) implies
(3.4)
where sc and si are consumption and investment shares in output and
As in other linear optimal control settings, expressions (3.1) and (3.2)
as functions of the state
can be solved to give optimal decisions
and the co-state (shadow price)
Further, given these
variables
(conditionally) optimal decisions, expressions (3.3) and (3.4) imply a firstorder dynamic system in and
(3.5)
where W is a 2×2 matrix and R and Q are 2×1 vectors. To compute the
solution to this difference equation and to examine its properties, we use
the decomposition W=PµP –1, where P is the matrix of characteristic vectors
of W and µ is a diagonal matrix with the characteristic roots on the diagonal.
Ordering the roots (µ1, µ2) in increasing absolute value, it can be shown that
15
When the utility function is additively separable, it follows that ξ cc=–1, ξ cl=ξ lc=0
and ξ ll=LD2(L)/D(L). When the utility function is multiplicatively separable, it
follows that ξ cc=-, ξ cl=LD(L)/v(L), ξ lc=1- and ξ ll =LD2(L)/D(L).
16
With the Cobb-Douglas assumption, it follows that ηA=[γ X-*(1-δK)]/γX, ηk=αηA and ηN=αηA.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE BASIC NEOCLASSIC MODEL
206
119
R.G.King et al., Production, growth and business cycles I
0<µ1<1<β*–1<µ2. The general solution to the difference equation for
and
is given by
specified initial conditions
(3.6)
Since Wt=PµtP–1 and the root µ2 exceeds (β *)–1>1, it follows that the system
is on an explosive path and thus violates the transversality condition for
There is a specific value of the initial shadow price
however,
arbitrary
that results in (3.6) satisfying the transversality condition (2.15). This
particular solution specifies the unique optimal (and competitive
and shadow prices
equilibrium) time path of capital accumulation
Given these optimal sequences, consumption
and effort
can be computed from (3.1) and (3.2). It is also direct to compute
variations in output, investment, real wages and real interest rates. For
in (3.4).
example, output variations are given by
With Cobb-Douglas production, real wages are proportional to labor
productivity, so that
In general, optimal decisions for consumption, capital, effort, etc. depend
As demonstrated in the technical appendix,
on the entire sequence
the time path of efficient capital accumulation may be written in the form
(3.7)
where 1 and 2 are complicated functions of the underlying parameters
of preferences and technology. The dynamics of capital accumulation
depend on the previous period’s capital stock with coefficient µ1. In
addition, with time-varying total factor productivity, the optimal solution
for capital accumulation depends on the current productivity level
and on the entire future time path of displacements to productivity
‘discounted’ by µ2.
3.2. Transition path dynamics
In order to provide a quantitative evaluation of the dynamic properties of
the neoclassical model we choose a set of parameters values that match the
average U.S. growth experience. The properties of the transition paths to
the steady state capital stock (k) can then be numerically investigated by
setting At=A for all t. In this case, the (approximately) optimal sequence of
transformed capital stocks described by (3.7) reduces to the first-order
with
Given an initial condition
difference equation
k0=K0/X0, the transformed economy’s capital stock approaches its steady
state value more quickly the closer µ1 is to zero. In addition, the variations
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
120
FOUNDATIONS
R.G.King et al., Production, growth and business cycles I
207
in consumption, investment, output, work effort, the real wage and the
real interest rate are determined according to linear relations:
(3.8)
where r is the steady state real interest rate, r=X/*–1. Except for πrk, the
π coefficients should be interpreted as the elasticities of the flow variables
with respect to deviations of the capital stock from its stationary value.
The transition paths of these flow variables, therefore, are simply scaled
versions of the capital stock’s transition path. In general, the values of µ1
and the π coefficients are complicated functions of the underlying
parameters of the model, i.e., α, σ, δ K, β and X.
3.2.1. A fixed labor experiment
Within the neoclassical model with fixed labor, variations in alter
substitution over time. Table 1 summarizes the quantitative effects of
varying on the adjustment parameter µ1 and the π coefficients.17 The
values of the underlying parameters assume that the time interval is a
quarter and are summarized in the table. Labor’s share α=0.58 is the
average ratio of total employee compensation to GNP for the period 1948
to 1986; X is one plus the common trend rate of growth of output,
consumption and investment, which is 1.6% per year in the post-war
era.18 The value for β*=X/(1+r) is chosen to yield a return to capital of
6.5% per annum, which is the average real return to equity from 1948 to
1981.19 Finally, the depreciation rate is set at 10% per annum, which
leads to a share of gross investment of 0.295.
In the fixed labor model, some of the π coefficients are invariant to .
The elasticities of output and real wages with respect to capital are simply
determined by πyk=πwk=(1–α) which is 0.42 in our table 1 example. The
value of πrk=ηk is also invariant to and takes the value –0.024. This
means that output and real wages move directly with capital and real
interest rates inversely with capital.
In the case of log utility (=1), the table shows that the adjustment
coefficient (µ1) is 0.966 which implies that one-half of any initial deviation
from the stationary state is worked off in 20 quarters or 5 years. If the
17
Tedious algebra shows that πck=[(1–α)–(γXµ1–(1–δK ))(k/y)]/sc and πik=[(γXµ1– (1δK))(k/y)]/si. It is direct that πNk=0 and πyk=(1-α). Finally, µ1 is the smaller root of the
quadratic equation µ2-[1/β *-scηK/ si+1]µ+1/β*.
18
Details of this computation and the data used are discussed in section 5.2.
19
Note that while β* is invariant with respect to under the assumption that
β*=γX/(1+r), β is not since
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
208
Half-life is defined by the solution to
rounded to the nearest integer.
THE BASIC NEOCLASSIC MODEL
a
R.G.King et al., Production, growth and business cycles I
Table 1
Effects of intertemporal substitution in consumption on near steady state dynamics (fixed labor model).
121
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
122
FOUNDATIONS
R.G.King et al., Production, growth and business cycles I
209
capital stock is initially below its steady state value, then investment is
above its steady state rate (πik=–0.176<0) and consumption is below its
steady state rate (πck=0.670>0).
Alternative values of change πck, πik and µ1 in intuitive ways. For
example, when is large the representative agent is less willing to substitute
intertemporally and thus desires very smooth consumption profiles. Hence,
there is little reaction of consumption to a shortfall in capital (πck small).
Consequently the adjustment to the steady state is slower (µ1 closer to 1.0)
than when =1.0. When is small, there is more willingness to substitute
consumption intertemporally and thus a given capital shortfall occasions
a larger reduction in consumption. There is thus a more rapid adjustment
of capital (µ1 further from 1.0) than with =1.
3.2.2. Varying work effort
We are also interested in the pattern of efficient variation in work effort
along the transition path, how the labor-leisure margin alters the speed of
capital stock adjustment (µ1) and the responses of the price and quantity
variables. To investigate these effects quantitatively, we reinstate labor as a
choice variable and suppose that the utility function has the simple form
u(c, L)=log(c)+llog(L). The parameter l is chosen so that stationary hours
are 0.20.20 Our choice of this value is based on the average percentage of
time devoted to market work in the U.S. during the period 1948–1986.21
The resulting value of µ1 is 0.953, implying a half-life of just under 14
quarters for deviations of the capital stock from its stationary level. This is
a slightly more rapid pace of adjustment than the comparable fixed labor
case with =1 in table 1, since work effort provides an additional margin
along which agents can respond. The values of the elasticities are πck=0.617,
πik=–0.629, πNk=–0.294, πyk=0.249, πwk=0.544 and πrk=–0.029. Transition
paths of the key variables are plotted in fig. 1. Starting from an initially
low capital stock, there is a sustained period in which output and
consumption are low, but rising, while work effort and investment are
high, but declining. Temporary variation in work effort is efficient even
though steady state hours are invariant to growth.
The economic mechanisms behind these transition paths are important.
The initially low capital stock has three implications for the representative
consumer in the transformed economy. First, non-human wealth is low
20
In our computations, we directly specify that N=0.20 in the linear expressions
(3.1) and (3.2), noting that logarithmic utility implies zero cross elasticities and unitary
own elasticities. This implicitly specifies the utility function parameter l.
21
This value is equal to the average work week as a fraction of total weekly hours
for the period 1948 to 1986.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE BASIC NEOCLASSIC MODEL
210
123
R.G.King et al., Production, growth and business cycles I
Fig. 1
relative to its stationary level. Second, the marginal product of labor (shadow
real wage) is low relative to the stationary level. Third, the marginal
product of capital (shadow real interest rate) is high relative to its stationary
level. The first and third factors induce the representative consumer to
work additional hours; the second factor exerts the opposite influence.
With the particular preferences and technology under study, the former
factors dominate, resulting in hours that are high—relative to the stationary
level—along the transition path from a low initial capital stock.
It is beyond the scope of this paper to undertake a detailed sensitivity
analysis of how the µ and coefficients change with parameters of the
environment. However, we have studied how the root µ1 depends on a
list of parameter values by computing an elasticity of µ1 with respect to
each parameter.22 The elasticities are quite small ranging from –0.11 for
labor’s share () to –0.001 for the rate of technological progress (X–1).23
22
We thank Adrian Pagan for pushing us to conduct these experiments.
The elasticity for steady state hours (N) is 0.003; for depreciation (K) is –0.03 for
the intertemporal elasticity of substitution ( ) is 0.03, and for the elasticity of the
marginal utility of leisure (LD2(L)/D(L)) is 0.003.
23
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
124
FOUNDATIONS
R.G.King et al., Production, growth and business cycles I
211
Our conclusion is that the speed of adjustment is not highly sensitive to
the choice of parameter values.
4. Real business cycles
This section follows the pioneering work of Kydland and Prescott (1982)
and Long and Plosser (1983) by incorporating uncertainty—in the form of
temporary productivity shocks—into the basic neoclassical model. Although
other aspects of the underlying economic environment are identical to
those of the preceding section, the business cycle analysis is in marked
contrast to the standard ‘growth theory’ analysis, in which time variation
in technology is taken to be smooth, deterministic and permanent.
4.1. Linear business cycle models
In principle, quantitative analyses of stochastic elements should follow Brock
and Mirman’s (1972) seminal analysis of the basic neoclassical model under
uncertainty. One would begin by postulating a specific stationary stochastic
process for technology shocks, calculate the equilibrium laws of motion for state
variables (the capital stock) and related optimal policy functions for controls
(consumption, investment and work effort). It would then be natural to interpret
observed business fluctuations in terms of the economy’s stationary distribution.
The principle barrier to the execution of this strategy is computational. The
equilibrium laws of motion for capital and for flows cannot be calculated exactly
for models of interest, but must be approximated with methods that are
computationally burdensome.24 Further-more, computational strategies for
approximate suboptimal equilibria are not well developed.
In our analysis we invoke certainty equivalence, employing a linear
systems perspective. Our use of certainty equivalence methods in the study
of real business cycles builds on prior work by Kydland and Prescott
(1982), but the details of our procedures are different.25 An advantage of
24
Examples include Sargent (1980) and Greenwood, Hercowitz and Huffman (1986).
Kydland and Prescott (1982) eliminate non-linearities in constraints (such as the
production function) by substituting resource constraints into the utility function and
taking a quadratic approximation to the resulting return function. We derive efficiency
conditions under certainty and approximate these to obtain linear decision rules. These two
procedures are equivalent for the class of models we consider when standard Taylor series
approximations are used with each procedure. The only substantive difference between our
approximation method and Kydland and Prescott’s is that while they search for an
approximation based on a likely range of variation of the different variables, we center our
linearizations on the steady state. According to Kydland and Prescott (1982, p. 1357, 11) this
difference in approximation techniques has little impact on their results. Our procedure
yields formulas that have a transparent economic interpretation and allows us to replicate
exactly the Long and Plosser (1983) closed form solution.
25
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE BASIC NEOCLASSIC MODEL
212
125
R.G.King et al., Production, growth and business cycles I
our method is that it is readily extended to the study of suboptimal dynamic
equilibria, as we show in our second essay. Nevertheless, a detailed analysis
of the overall accuracy of these approximation methods in a business cycle
context remains to be undertaken.
For the basic neoclassical model, our strategy works as follows. We develop
approximate solutions for capital and other variables near the stationary
point of the transformed economy as in the previous section. Then, working
from a certainty equivalence perspective, we posit a particular stochastic
process for
and replace the sequence
with its conditional
expectation given information available at t. In particular, suppose that
follows a first-order autoregressive process with parameter ρ. Then, given
(3.7), the state dynamics are given by the linear system
(4.1)
where
and
is the state vector.
Additional linear equations specify how consumption, work effort,
investment, shadow prices and output depend on the state variables st.
Let the vector
be a vector of controls and other flow
variables of interest. Then the linear equations relating flows to states are
(4.2)
where the coefficients are determined, as in section 3, by requiring that the
shadow prices and elements of zt satisfy the linearized first-order conditions.26
26
This state space formulation (4.1) and (4.2) can be solved to obtain the vector autoregressivemoving average (ARMA) representation of the endogenous variables z. In the basic neoclassical
model with persistent technology shocks (≠0), each element of zt is ARMA (2, 1) with
common autoregressive but different moving average polynomials. Following Zellner and
Palm (1974) and Orcutt (1948), the evolution of states can be expressed as follows
where B is the backshift operator, det(I–MB) is the determinant of the 2×2 matrix defined by
I–MB, and adj(I–MB) is the adjoint of I–MB. From inspection of (4.1) it is clear that, for ≠0,
the determinant of (I–MB) is a second-order polynomial (1-µ1B)(1-B). There are moving
average terms of at most order 1 in adj(I–MB). Further, since zt=st, the elements of zt inherit
the ARMA (2, 1) structure of the state vector. The relatively simple ARMA structure of the
individual elements of z is a result of the dimensionality of the state vector. In a model with
many state variables the order of the polynomial det(I–MB) could become quite large, implying
more complex ARMA representations for the elements of z.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
126
FOUNDATIONS
R.G.King et al., Production, growth and business cycles I
213
This formulation facilitates computation of (i) impulse response functions
for the system and (ii) population moments of the joint (zt, st) process.
Impulse responses. Impulse response functions provide information on the
system’s average conditional response to a technology shock at date t,
given the posited stochastic process. The response of the system in period
t+k to a technology impulse at date t+1 is
Population moments. Population moments provide additional, unconditional
properties of the time series generated by the model economy. We stress
that although there is a single shock to the economic model under study,
the dynamic character of the model means that the unconditional time
series will, in general, not be perfectly correlated. The linear character of
the system implies that it is relatively straightforward to calculate population
moments. For example, given the variance-covariance matrix of the states,
it is easy to calculate the autocovariance of z at lag j,
In our analysis below, we will be concerned with
how these properties of the model change as we alter parameters of
preferences and technology.
4.2. Alternative parameterizations of the basic neoclassical model.
We explore four alternative parameterizations of the basic neoclassical
model, obtained by varying certain aspects of preferences and technology.
Though far from exhaustive, these parameterizations shed some light on
important aspects of neoclassical models. Table 2 summarizes the parameter
values that are employed in our four versions of the neoclassical model.
Throughout, as in table 1, we use production parameter values for labor’s
share as α=0.58 and the growth of exogenous technical progress as (γX–1)
=0.004 per quarter. In all specifications, we take the momentary utility
function to be of the additively separable form, u(c, L)=log(c)+θl␷(L). This
specification implies zero cross-elasticities (ξlc=ξcl=0) and unitary elasticity
in consumption (␴=–ξcc=1), while leaving the elasticity of the marginal
utility of leisure with respect to leisure (ξll) as a parameter to be specified.
The parameter θl in all parameterizations is adjusted so as to yield a steady
state value for N equal to 0.20, the average time devoted to market work
in the U.S. during the period 1948–1986. In all of these formulations, the
values of ␴, γX and β combine to yield a steady state real interest rate of
6.5% per annum.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
214
Table 2
Alternative model parameterizations.
THE BASIC NEOCLASSIC MODEL
R.G.King et al., Production, growth and business cycles I
127
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
128
FOUNDATIONS
R.G.King et al., Production, growth and business cycles I
215
Our point of departure is the parameterization of Long and Plosser (1983).
The key features of this specification are additively separable, logarithmic
preferences, a Cobb-Douglas production function and 100% depreciation.
This specification is instructive because there is an exact closed-form solution
that enables us to establish a benchmark for judging our approximation
methods. The second specification alters the Long-Plosser formulation by
assuming less than 100% depreciation. This alteration is sufficient to obtain
stochastic properties for key variables that are more compatible with common
views of economic fluctuations. We refer to this case as the ‘baseline’ model—
it is closely related to the divisible labor economy studied by Hansen (1985).27
The next two experiments consider some perturbations of the elasticity of
labor supply. The third parameterization uses an ‘upper bound’ labor supply
elasticity from the panel data studies reviewed by Pencavel (1986). This
elasticity is ten times smaller than that imposed by the logarithmic preferences
of the baseline mode.28 The fourth parameterization illustrates the
consequences of infinite intertemporal substitutability of leisure or,
equivalently, the indivisibility of individual labor supply decisions stressed
by Rogerson (1988) and Hansen (1985).
4.3. Quantitative linear business cycle models
The reference point for our discussion is table 3, which summarizes the linear
systems representation given in eqs. (4.1) and (4.2). That is, table 3 provides the
coefficients, µ1, , πkA of the matrix M and the coefficients of the Π matrix under
two assumptions about persistence of technology shocks (=0 and =0.9).
Long-Plosser with complete depreciation. Applying the exact solutions found in
Long and Plosser (1983), the capital stock for this parameterization evolves
according to the stochastic difference equation,
27
There are at least three differences between our methodology and that employed by
Hansen (1985) which make our results not directly comparable. First, we use a different
linearization technique, as discussed above. Second, we compute the population movements
rather than estimate them through Monte Carlo simulation. Third, we do not filter the
series with the Hodrick and Prescott (1980) filter. See footnote 31 for a discussion of
differences in parameter values and of the effects of the Hodrick and Prescott filter.
28
For preferences separable in consumption and leisure, the elasticity of labor supply is
(1–1/N)/ξ ll, where N is the steady state fraction of time devoted to work. Thus if the
elasticity of labor supply is 0.4 and N=0.20, then ξ ll=-10.0.
We are reluctant to adopt this economy as our benchmark given the difficulty in
interpreting the disparity between the elasticity of labor supply of women and men in the
context of our representative agent economy. Furthermore, Rogerson (1988) has
demonstrated that, in the presence of indivisibility in individual labor supply decisions, an
economy with finite elasticity of labor supply may behave as if this elasticity were infinite.
Hence, our fourth parameterization has preferences consistent with an infinite elasticity
of labor supply (ξll=0).
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
216
Table 3
Parameter values of the linear system (4.1)–(4.2).
THE BASIC NEOCLASSIC MODEL
R.G.King et al., Production, growth and business cycles I
129
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
130
FOUNDATIONS
R.G.King et al., Production, growth and business cycles I
217
(4.3)
which indicates that in our approximation it should be the case that µ1=
(1–) and π kA =1.0. As emphasized by Long and Plosser, (4.3) illustrates
that even without long-lived commodities, capitalistic production enables
agents to propagate purely transitory productivity shocks forward in time
in keeping with their preferences for smooth consumption.
The solutions of Long and Plosser also imply that there are simple loglinear relations for the flow variables
(4.4)
(4.5)
In percent deviations from steady state, output, consumption, and
investment all share the stochastic structure of the capital stock. Work
effort, on the other hand, is constant (i.e.
,). With work effort constant,
real wages (proportional to output per man hour) move just like output.
With =1, interest rates are equal to the expected change in consumption
Thus, in terms of (4.2), yk=
ck=
ik=
wk=(1–),
yA = cA = iA = Nk =1, and Nk = NA =0 . Finally, rk =- (1– ) and
Turning to the approximate solutions reported in table 3, we see that
these match the exact solutions (4.3)–(4.5) for the parameter values in table
2. For example, with =0.58, the coefficient µ1=(1–)=0.42 as required by
eq. (4.1) above. Further, we see that there are two special features of this
parameterization previously noted by Long and Plosser in their
consideration of multi-sector, log-linear business cycle models. First, the
solution involves no influence of expected future technological conditions
on the properties of the endogenous variables. This conclusion follows
from the observation that the linear systems coefficients linking quantity
variables to technology (πkA, πcA, πNA, etc.) are invariant to the persistence
() in the technology shock process. Second, the relation between work
effort and the state variables (πNk and πNA) indicates that the approximation
method preserves the other special implications of complete depreciation,
namely that effort is unresponsive to the state of the economy (πNA=πNk=0).
Fundamentally, each of these invariance results reflects a special
balancing of income and substitution effects. For example, more favorable
exert two offsetting effects on
technology conditions
accumulation: (i) an income effect (since there will be more outputs at
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE BASIC NEOCLASSIC MODEL
218
131
R.G.King et al., Production, growth and business cycles I
given levels of capital input) that operates to lower saving and capital
accumulation and (ii) a substitution effect (arising from an increased
marginal reward to accumulation) that operates to raise saving. With
complete depreciation and logarithmic utility, income and substitution
effects exactly offset.
With respect to real interest rates, the complete depreciation model
alters the model’s
also helps indicate how serial correlation in
implications. The coefficient rA=(–a), so that with =0 diminishing
returns predominates and an impulse to A lowers the rate of return. But
with high persistence (>), interest rates rise due to the shift up in the
future marginal reward to investment.
Long-Plosser with realistic depreciation. Adopting a more realistic depreciation
rate (K=0.025 or 10% per year) dramatically alters the properties of the
basic neoclassical model. The adjustment parameter µ1 rises from 0.42 to
0.953, indicating that the capital stock adjusts more slowly. Second, kA
falls from 1.0 to 0.166 when =0 and is no longer invariant to serial
correlation properties of
These responses can be explained in terms of the basic economics of
lowering the depreciation rate. First, when there is a lower depreciation
rate, it follows that there is a higher steady state capital stock and a lower
output-capital ratio. As K goes from 1.0 to 0.025, y/k falls from 2.4 to 0.10.
This suggests a substantial decline in the elasticity kA. Second, the change
in µ1 and the sensitivity of kA to reflect implications that K has for the
relative importance of wealth and intertemporal substitution effects. With
lower depreciation, the intertemporal technology—linking consumption today
and consumption tomorrow—becomes more linear near the stationary point.29
This means that the representative agent faces less sharply diminishing
returns in intertemporal production possibilities and will choose a temporally
smooth consumption profile that requires more gradual elimination of
deviations of the capital stock from its stationary level (µ1 rises from 0.42
when K=1 to 0.953 when K=0.025). The depreciation rate also impinges
on the relative importance of substitution and wealth effects associated with
for j >0). In particular, the dominance of
future shifts in technology (
the wealth effect is indicated by a comparison of purely temporary (=0)
with more persistent technology shocks. Capital accumulation is less
responsive to technological conditions when the shocks are more persistent
(i.e., kA falls from 0.166 to 0.137 when rises from 0 to 0.9). For the same
reason, more persistent technology shocks imply that consumption is more
responsive (
cA=0.108 when =0 and cA=0.298 when =0.9) and
29
There is a marked decline in the elasticity of the gross marginal product of capital
schedule, AD1F(k, N)+(1–K), with respect to capital. It falls from –k=0.58 to 0.023.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
132
FOUNDATIONS
R.G.King et al., Production, growth and business cycles I
219
investment is less responsive (
iA=5.747 when =0 and iA=4.733 when
=0.9). The persistence of shocks also has implications for the response of
relative prices to technology shifts. Real wages respond more elastically,
since there is a smaller variation in effort when shifts are more permanent.
As in the model with complete depreciation, real interest rates respond
positively to technology shifts when these are highly persistent.
Altering the character of intertemporal tradeoffs also has implications for
labor supply via intertemporal substitution channels. When technology shifts
are purely temporary (=0), a one percent change in total factor productivity
calls forth a 1.33 percent change in hours. This impact is attenuated, but not
eliminated, when shifts in technology are more persistent (
NA=1.05 when
=0.9). The nature of these intertemporal substitution responses is perhaps
best illustrated by examining impulse response functions, which are derived
from the coefficients presented in table 3. The two parts of fig. 2 contain
impulse responses under our alternative assumptions about the persistence of
shocks. In panel A, when technology shifts are purely temporary, intertemporal
substitution in leisure is very evident. In the initial period, with positive one
percent technology shock, there is a major expansion of work effort. The
initial period output response is more than one-for-one with
because of the expansion in work effort. The bulk of the output increase goes
into investment with a smaller percentage change in consumption.
In subsequent periods, after the direct effect of the technology shift has
dissipated, the only heritage is a capital stock higher than its steady state value.
The change in the capital stock induced by the initial period technology shock
is ‘worked off’ via a combination of increased consumption and reduced effort.
The impacts on output are smaller, in percentage terms, than the impacts on
consumption or capital, because the transition path back toward the stationary
point is associated with negative net investment and negative response of effort.
This means that the response function after one period in fig. 2, panel A, is
determined by the internal transition dynamics given in fig. 1. The only
difference is that in fig. 2 the experiment is a positive increment to the capital
stock of 0.166 instead of the negative increment of –1.0. in fig. 1.
In panel B of fig. 2, when technology shifts are more persistent, the
and endogenous
impulse responses involve a combination of exogenous
There is now a protracted period in which technology shocks
dynamics
serve to introduce positive comovements of hours, output, consumption
and investment. The magnitudes of these responses are altered by the fact
that agents understand the persistent character of technological shifts. In
comparison with the case where technology shifts are purely temporary,
consumption is more responsive to while effort is less.
Other labor supply elasticities. First, when we restrict preferences to be consistent
with an ‘upper bound’ labor supply elasticity of 0.4 for prime age males
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE BASIC NEOCLASSIC MODEL
220
R.G.King et al., Production, growth and business cycles I
Fig. 2
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
133
134
FOUNDATIONS
R.G.King et al., Production, growth and business cycles I
221
reported by Pencavel (1986), we obtain an economy whose dynamics are
broadly similar to those of the baseline model except for the amplitude of
response to technology shocks. In the case of purely temporary production
shocks (=0), the elasticity of response of labor to shocks in technology
(
NA) is 0.317, roughly one fourth of the value of NA for the baseline
model. This reduced variability in labor is accompanied by smaller
variability in consumption and investment.30
Second, when the labor supply elasticity is infinite, we have an economy
that is the mirror image of the previous one in terms of amplitude of response to
shocks. In the case of purely temporary shocks, the values of cA and iA are
roughly 1.2 times those of the baseline model, while NA is fifty percent higher.
5. Implications for time series
This section develops some of the predictions that the basic neoclassical
model makes about economic time series when it is driven by a single
technology shock. Using the model’s organizing principles, we also present
some summary statistics for post-war quarterly U.S. time series.
5.1. Variability of components of output
A major feature of economic fluctuations is the differential variability in
the use of inputs (labor and capital) and in the components of output
(consumption and investment). Table 4 presents some selected population
moments for the four alternative parameterizations that summarize the
models’ implications for relative variability.
The specification with complete depreciation has implications that are
readily traced to the simple structure of (4.3) and (4.4). First, output,
consumption and investment have identical variances. Second, with
complete depreciation, investment and capital differ only in terms of
timing, so that capital and output are equally variable.
When realistic depreciation is imposed (K=0.025), differences in the relative
variability of the components of output are introduced. Further, these
implications depend on the stochastic process for the technology shifts, since
the moments of time series depend on the linear system coefficients reported
in table 3 (which are dependent on the persistence parameter ). With purely
temporary shocks, consumption is much less variable than output (about two
tenths as variable). and investment is far more variable (more than three
times as variable). Labor input is much more variable than consumption
and about three fourths as variable as output.
30
A productivity shock induces intertemporal substitution of leisure by raising the
productivity of current versus future labor and intratemporal substitution by
increasing the opportunity cost of leisure in terms of consumption. Both the elasticity
of intertemporal substitution of leisure and the elasticity of intratemporal substitution
are smaller in this economy than in the baseline model. The reduction in the degree
of substitution contributes to a reduced variability of consumption.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
222
Table 4
Selected population moments for four alternative parameterizations.
THE BASIC NEOCLASSIC MODEL
R.G.King et al., Production, growth and business cycles I
135
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
136
FOUNDATIONS
R.G.King et al., Production, growth and business cycles I
223
When shifts in technology become more persistent (=0.9), there are
important changes in implications for relative variabilities. Consumption
is now six tenths as volatile as output, which accords with the permanent
income perspective and with the altered linear systems coefficients
discussed previously. Labor input is less than half as volatile as output,
which fundamentally reflects diminished desirability of intertemporal
substitution of effort with more persistent shocks.31
Alterations in the labor supply elasticity exert predictable effects on
relative variability of labor input and output, while having relatively minor
implications for the relative variability of the components of output.
Relative to the baseline model, the reduction in labor supply elasticity to
the level suggested by the panel data studies results in a decline of the
variability of labor both in absolute terms and in relation to the variability
of output. The relative volatility of the labor input in terms of output
implied by the model is 0.27, roughly half of the standard deviation of
hours relative to detrended output in the U.S. for the period 1948–1986.32
In table 5 we present some additional time series implications of our
and exhibit
baseline neoclassical model. One notable feature is that
almost no serial correlation in the absence of serially correlated technology
shocks. This is not true for consumption, wages or interest rates, however,
which are smoother and correlated with lagged values of output.
31
The baseline model is structurally identical to the divisible labor economy studied by
Hansen (1985). It differs, however, in values assigned to parameters. In our notation, Hansen’s
economy involves =0.64, β*=0.99, X=1.00; N=0.33 and K=0.025. These alternative
parameter values have implications for the moments reported in tables 4 and 5. Using a
persistence parameter =0.90, the model’s relative volatility measures (standard deviations of
variables relative to standard deviation of output) are as follows: consumption (0.62), investment
(2.67) and hours (0.41). Basically, relative to table 4 these results reflect the decline in labor
supply elasticity implied by N=1/3 rather than N=1/5. The contemporaneous correlations
with output are as follows: consumption (0.81), investment (0.92) and hours (0.81). If we filter
the population moments with the Hodrick-Prescott (HP) filter, then the relative variabilities
and correlations are altered. For consumption these are (0.25) and (0.80), respectively, for
investment they are (3.36) and (0.99) and for hours they are (0.55) and (0.98). These alterations
occur because the effect of the HP filter is to give less weight to low frequencies, downplaying
persistent but transient aspects of the series in question. [See the graph of the transfer function
of the HP filter in Singleton (1988).] For example, the correlation of output at the yearly
interval (lag 4) is 0.72 in the unfiltered Hansen parameterization and it is 0.08 in the filtered
version. It is this sensitivity of results to filtering that makes us hesitant to undertake detailed
comparisons with results reported by Hansen.
32
The inability of the model to generate a sufficiently high variation in labor when the
elasticity of labor supply is restricted to be consistent with panel data studies has stimulated
several extensions to the basic neoclassical model. Kydland (1984) demonstrates that introducing
agent heterogeneity in the model can increase the relative volatility of the average number of
hours worked with respect to the volatility of labor productivity. Rogerson (1988) establishes
that, in the presence of indivisibility in individual labor supply, an economy with finite
elasticity of labor supply behaves as if it had an infinite elasticity of labor supply. This
motivates our interest in the fourth parameterization. As Hansen (1985), we find that in this
economy labor is too volatile relative to output.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
224
Table 5
Population moments: Baseline model.
THE BASIC NEOCLASSIC MODEL
R.G.King et al., Production, growth and business cycles I
137
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
138
FOUNDATIONS
R.G.King et al., Production, growth and business cycles I
225
5.2. Some empirical issues and observations
Since the early part of this century, with the NBER studies of business
cycles and economic growth under the leadership of Wesley Mitchell
and Simon Kuznets, it has become commonplace for macroeconomic
researchers to design models to replicate the principal features of the
business cycles isolated by the NBER researchers. More recently, the
development of statistical and computing technology has led individual
researchers to define analogous sets of ‘stylized facts’ about economic
fluctuations that models are then designed to emulate.
Our perspective is that the development of stylized facts outside of a
circumscribed class of dynamic models is difficult at best.33 First, models
suggest how to organize time series. Further, it is frequently the case that
stylized facts are sensitive to the methods of detrending or prefiltering. In
this investigation we take the perspective that the basic neoclassical model
has implications for untransformed macroeconomic data and not some
arbitrary or prespecified transformation or component that is defined outside
the context of the model [cf. Hicks (1965 p. 4)]. Although we do not
perform formal statistical tests of model adequacy, the manner in which
we proceed with data analysis is dictated by the models under study.
We have considered deterministic labor augmenting technological change
that grows at a constant proportionate rate as the source of sustained growth
(trend). The neoclassical model then predicts that all quantity variables (with
the exception of work effort) grow at the same rate X. The non-deterministic
components of consumption, output and investment are then
(5.1)
where y, c and i are the steady state values in the transformed economy.
Labor augmenting technical progress, log(Xt), can be expressed as the
simple linear trend
(5.2)
Thus, in the language of Nelson and Plosser (1982), the implied time series
are trend stationary. Moreover, they possess a common deterministic trend.
Therefore, the model instructs us to consider deviations of the log levels of
GNP, consumption and investment from a common linear trend as empirical
counterparts to , and Work effort, on the other hand, possess no
trend and, thus, is simply deviation of the log of hours from its mean.
33
See also Koopmans (1947) and Singleton (1988).
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE BASIC NEOCLASSIC MODEL
226
139
R.G.King et al., Production, growth and business cycles I
In order to provide some perspective on the models’ properties, we summarize
some of the corresponding sample moments of the U.S. time series. The series we
consider are the quarterly per capita values of real GNP, consumption of nondurables and services (CNS), gross fixed investment (GFI) and average weekly
hours per capita.34 Following the structure (5.1) and (5.2), we detrend the log
levels of each of the first three series by computing deviations from a common
estimated linear time trend. The estimated common trend, which corresponds
is 0.4% per quarter.35 The real wage is the
to an estimate of
gross average hourly earnings of production or non-supervisory workers on
non-agricultural payrolls. We chose not to study interest rates because of the
well-known difficulties of obtaining measures of expected real interest rates.
and are presented in fig.
Plots of our empirical counterparts to
3. Their properties are summarized in table 6 in a manner analogous to
the summary of the baseline model in table 5. Our sample period is the
first quarter of 1948 (1948.1) to the fourth quarter of 1986 (1986.4).
Deviations of output from the common deterministic trend, which are
plotted as a benchmark in each of the panels in fig. 3, have a standard
deviation of 5.6% and range in value from –13.0% to 10%. The sample
autocorrelations in table 6 indicate substantial persistence, suggesting that
there may be a non-stationary component to the series not eliminated by
removing a common deterministic trend.
The panels A and B show empirical counterparts to and
plotted
against the reference variable . Consumption and investment are highly
correlated with output. Table 6 reports estimated correlation coefficients
of 0.85 for consumption and 0.60 for investment over the 1948.1–1986.4
sample period. Consumption is less volatile than output, with a sample
standard deviation of 3.9% (versus 5.6% for output) and a sample range
of –7.8% to 7.4%. Investment is more volatile than output, with a sample
standard deviation of 7.6% and sample range of –20.7% to 16.3%. Further,
the autocorrelation statistics in table 6 indicate substantial serial correlation
in both consumption and investment.
Panel C of fig. 3 contains a plot of the empirical counterpart of per capita
hours as well as that of output. This labor input measure has a standard
deviation of 3.0%, with a maximum value of 6.5% and a minimum value of
34
All series are taken from the CITIBASE database. GNP, CNS and GFI are
quarterly values. Population (P) is the total civilian non-institutional population 16
years of age and older. Employment (E) is total workers employed as taken from the
Household Survey, Bureau of Labor Statistics. Average weekly hours of all workers
(H) is also from the Household Survey. Average hours per capita is then calculated as
E·H/P and averaged for the quarter. The wage rate is gross average hourly earnings
of production workers.
35
This is the source of the estimate of X we use to parameterize the basic model in
section 3. We choose not to impose the common trend assumption on wage rates
because it involves a specific assumption about market structure.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
140
FOUNDATIONS
R.G.King et al., Production, growth and business cycles I
Fig. 3
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
227
THE BASIC NEOCLASSIC MODEL
228
R.G.King et al., Production, growth and business cycles I
Fig. 3 (continued)
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
141
142
b
All variables are taken from the National Income Accounts.
Relative standard deviation of z is
229
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
FOUNDATIONS
a
R.G. King et al., Production, growth and business cycles I
Table 6
Sample moments: Quarterly U.S. data, 1948.1–1986.4.a
THE BASIC NEOCLASSIC MODEL
230
143
R.G.King et al., Production, growth and business cycles I
–6.2% over the post-war period. The correlation between output and hours
reported in table 6 is essentially zero! Inspection of the plot, however, appears
to suggest that this relation is closer if one visually corrects for the periods
in which output is on average high or low. In fact, if one splits the sample
into subperiods of approximately 5 years each, the correlation between output
and hours is never less than 0.30 and averages 0.77. Thus, when we permit
the sample mean to vary (which is what looking at subperiods effectively
does), the correlation between hours and output appears much higher.36 It
is important to stress that there is no theoretical justification for looking at
data in subperiods. The basic neoclassical model that we have been discussing
has a single source of low frequency variation (the deterministic trend in
labor productivity) which has been removed from the time series under
study. The sensitivity of these results to the sample period suggests the
possibility of a low frequency component not removed by the deterministic
trend. This is consistent with the highly persistent autocorrelation structure
of output noted above.
The practice of removing low frequency variation in economic data plays
an important role in empirical research on business fluctuations. NBER
business cycle research has generally followed Mitchell’s division of time
series into cyclical episodes, removing separate cycle averages for individual
series. Our belief is that this methodology is likely to remove important low
frequency aspects of the relations between time series, in a manner broadly
similar to the computation of correlations over subperiods. Most modern
empirical analyses of cyclical interactions have also followed the practice of
removing low frequency components from actual and model-generated time
series.37 Studying the impact of such low frequency filtering on economic
time series generated by our baseline model, King and Rebelo (1987) find
that there are major distortions in the picture of economic mechanisms
presented by low frequency filtering. Among these are two that are
particularly relevant to the labor-output relation. First, in the theoretical
economy analyzed by King and Rebelo, application of a low frequency
filter raises the correlation between output and labor input. Second, a low
frequency filter dramatically reduces the correlation between output and
capital.
Panel D of fig. 3 contains a plot of our empirical measure of While
the correlation with output is positive (0.76), it is not as strong as predicted
36
The subperiod results for the other variables are qualitatively similar to the overall
sample. We have also explored the use of another hours series to insure that this finding was
not an artifact of our data. Using an adjusted hours series developed by Hansen (1985),
which covers only the 1955.3 to 1984.1 period, the correlation is 0.28 compared to 0.48 for
our series for the same period. Breaking this shorter sample into subperiods also yields higher
correlations than those for the overall period for the Hansen data.
37
For example, Kydland and Prescott (1982) filter both the data and the output of their
model using a filter proposed by Hodrick and Prescott (1980). Hansen (1985) follows this
practice as well.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
144
FOUNDATIONS
R.G.King et al., Production, growth and business cycles I
231
by the model. Moreover, the positive correlation seems to arise primarily
from the association at lower frequencies.
There are two main conclusions we draw from this cursory view of the
data. The first, and most important, is that the one sector neoclassical
model that we use as our baseline specification is not capable of generating
the degree of persistence we see in the data without introducing substantial
serial correlation into the technology shocks. The second is that the data
suggest the related possibility of a low frequency component not captured
by the deterministic trend. This motivates our interest in models with
stochastic growth in the companion essay.
6. Conclusions
This paper has summarized the growth and business cycle implications of
the basic neoclassical model. When driven by exogenous technical change
at constant rates, the model possesses a steady state growth path under
some restrictions on preferences for consumption and leisure. Although
these restrictions imply that labor effort is constant in the steady state,
they do not imply that effort is constant along transition paths of capital
accumulation or in response to temporary technology shocks. Rather, the
intertemporal substitution made feasible by capital accumulation applies
to both consumption and effort in general equilibrium.
When driven by highly persistent technology shocks, the basic
neoclassical model is capable of replicating some stylized facts of economic
fluctuations. First, the model generates procyclical employment, consumption
and investment. Second, the model generates the observed rankings of
relative volatility in investment, output and consumption. But along other
dimensions, the basic model seems less satisfactory. In particular, the principle
serial correlation in output—one notable feature of economic fluctuations—
derives mainly from the persistence of technology shocks. On another level,
as McCallum (1987) notes, the model abstains from discussing implications
of government and the heterogeneity of economic agents.
Perhaps the most intriguing possibility raised by the basic model is that
economic fluctuations are just a manifestation of the process of stochastic growth.
In the companion essay, we discuss current research into this possibility, along
with issues concerning the introduction of government and heterogeneity.
References
Barro, R. and R.King, 1984, Time separable preferences and intertemporal substitution
models of business cycles, Quarterly Journal of Economics 99, 817–839.
Becker, R., J.Boyd III and C.Foias, 1986, The existence of Ramsey equilibrium, Working
paper (University of Rochester, Rochester, NY).
Boyd III, J., 1986, Recursive utility and the Ramsey problem, Working paper no. 60 (Center
for Economic Research, University of Rochester, Rochester, NY).
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE BASIC NEOCLASSIC MODEL
232
145
R.G.King et al., Production, growth and business cycles I
Brock, W. and L.Mirman, 1972, Optimal economic growth and uncertainty: The discounted
case, Journal of Economic Theory 4, 479–513.
Cass, D., 1965, Optimum growth in an aggregative model of capital accumulation, Review
of Economic Studies 32, 223–240.
Debreu, G., 1954, Valuation equilibrium and Pareto optimum. Proceedings of the National
Academy of Sciences of the U.S.A. 38, 886–893.
Greenwood, J., Z.Hercowitz and G.Huffman, 1986, Investment, capacity utilization and the
real business cycle, Working paper (University of Western Ontario, London, Ont.).
Hansen, G., 1985, Indivisible labor and the business cycle, Journal of Monetary Economics
16, 309–327.
Hicks, J., 1965, Capital and growth (Oxford University Press, New York).
Hodrick, R. and E.Prescott, 1980, Post-war U.S. business cycles: An empirical investigation.
Working paper (Carnegie-Mellon University, Pittsburgh, PA).
King, R., 1987, Business cycles and economic growth, Lecture notes on macroeconomics
(University of Rochester, Rochester, NY).
King, R. and S.Rebelo, 1987, Low frequency filtering and real business cycles, in progress
(University of Rochester, Rochester, NY).
Koopmans, T., 1947, Measurement without theory, Review of Economics and Statistics 29,
161–172.
Koopmans, T., 1965, On the concept of optimal economic growth, in: The econometric
approach to development planning (Rand-McNally, Chicago, IL).
Kydland, F., 1984, Labor force heterogeneity and the business cycle, Carnegie-Rochester
Conference Series on Public Policy 21, 173–208.
Kydland, F. and E.Prescott, 1982, Time to build and aggregate fluctuations, Econometrica
50, 1345–1370.
Long, J. and C.Plosser, 1983, Real business cycles, Journal of Political Economy 91, 1345–
1370.
McCallum, B.T., 1987, Real business cycles, Unpublished manuscript (Carnegie-Mellon
University, Pittsburgh, PA).
Nelson, C. and C.Plosser, 1982, Trends and random walks in macroeconomic time series:
Some evidence and implications, Journal of Monetary Economics 10, 139–167.
Orcutt, G., 1948, A study of the autoregressive nature of the time series used for Tinbergen’s
model of the economic system of the United States, 1919–1932, Journal of the Royal
Statistical Society B 10, 1–45.
Pencavel, J. 1986, Labor supply of men: A survey, in: Orley Ashenfelter and Richard
Layard, eds., Handbook of labor economics (North-Holland, Amsterdam).
Phelps, E., 1966, Golden rules of economic growth (Norton, New York).
Prescott, E., 1986, Theory ahead of business cycles measurement, Carnegie-Rochester
Conference Series on Public Policy 25, 11–66.
Prescott, E. and R.Lucas, 1972, A note on price systems in infinite dimensional space,
International Economic Review 13, 416–422.
Rogerson, R., 1988, Indivisible labor, lotteries and equilibrium, Journal of Monetary
Economics 21, 3–16.
Romer, P. and T.Shinotsuka, 1987, The Kuhn-Tucker theorem implies the transversality
condition at infinity, Unpublished paper (University of Rochester, Rochester, NY).
Sargent, T., 1980, Tobin’s ‘q’ and the rate of investment in general equilibrium, CarnegieRochester Conference Series on Public Policy 12, 107–154.
Singleton, K., 1988, Econometric issues in the analysis of equilibrium business cycle models,
Journal of Monetary Economics, this issue.
Solow, R., 1956, A contribution to the theory of economic growth, Quarterly Journal of
Economics 70, 65–94.
Swan, T., 1963, On golden ages and production functions, in: Kenneth Berril, ed., Economic
development with special references to southeast Asia (Macmillan, London).
Weitzman, M., 1973, Duality theory for infinite time horizon convex models, Management
Science 19, 738–789.
Zellner, A. and F.Palm, 1974, Time series analysis and simultaneous equations models,
Journal of Econometrics 2, 17–54.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
Part III
Some extensions
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
CHAPTER 8
149
Journal of Monetary Economics 16 (1985) 309–327. North-Holland
INDIVISIBLE LABOR AND THE BUSINESS CYCLE
Gary D.HANSEN*
University of California, Santa Barbara, CA 93106, USA
A growth model with shocks to technology is studied. Labor is indivisible, so all variability
in hours worked is due to fluctuations in the number employed. We find that, unlike previous
equilibrium models of the business cycle, this economy displays large fluctuations in hours
worked and relatively small fluctuations in productivity. This finding is independent of
individuals’ willingness to substitute leisure across time. This and other findings are the result
of studying and comparing summary statistics describing this economy, an economy with
divisible labor, and post-war U.S. time series.
1. Introduction
Equilibrium theories of the business cycle, such as Kydland and Prescott (1982)
or Lucas (1977), have been criticized for failing to account for some important
labor market phenomena. These include the existence of unemployed workers,
fluctuations in the rate of unemployment, and the observation that fluctuations
in hours worked are large relative to productivity fluctuations. Equilibrium
models have also been criticized for depending too heavily on the willingness
of individuals to substitute leisure across time in response to wage or interest
rate changes when accounting for the last observation. This criticism is based
at least partially on the fact that micro studies using panel data on hours
worked by individuals have not detected the intertemporal substitution
necessary to explain the large aggregate fluctuations in hours worked [see
Ashenfelter (1984)].
In this paper, a simple one-sector stochastic growth model with shocks to
technology is constructed in which there is high variability in the number
employed and total hours worked even though individuals are relatively
unwilling to substitute leisure across time. The model differs from similar
models, such as Kydland and Prescott (1982), in that a non-convexity
* This paper is part of my doctoral dissertation written while a student at the University
of Minnesota. I have benefited from conversations with many people including Robert King,
Thomas Sargent, Christopher Sims, Neil Wallace, Sumru Altug, Patrick Kehoe, Ramon
Marimon, Ian Bain, and Rody Manuelli. I owe my greatest debt, however, to my advisor,
Edward Prescott. I wish to also acknowledge the Federal Reserve Bank of Minneapolis which
has provided support for this research. All errors, of course, are mine.
0304–3923/85/$3.30 © 1985, Elsevier Science Publishers B.V. (North-Holland)
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
150
310
SOME EXTENSIONS
G.D.Hansen, Indivisible labor and the business cycle
(indivisible labor) is introduced. Indivisible labor is modeled by assuming that
individuals can either work some given positive number of hours or not at
all—they are unable to work an intermediate number of hours. This assumption
is motivated by the observation that most people either work full time or not
at all. Therefore, in my model, fluctuations in aggregate hours are the result of
individuals entering and leaving employment rather than continuously employed
individuals adjusting the number of hours worked, as in previous equilibrium
models. This is consistent with an important feature of U.S. post-war data:
most fluctuation in aggregate hours worked is due to fluctuation in the number
employed as opposed to fluctuation in hours per employed worker. This is a
fact that previous equilibrium theories have not tried to account for.1
Existing equilibrium models have also failed to account for large fluctuations
in hours worked along with relatively small fluctuations in productivity (or the
real wage). Prescott (1983), for example, finds that for quarterly U.S. time
series, hours worked fluctuates about twice as much (in percentage terms) as
productivity. In this paper it is shown that an economy with indivisible labor
exhibits very large fluctuations in hours worked relative to productivity. This
stands in marked contrast to an otherwise identical economy that lacks this
non-convexity. In this economy hours worked fluctuates about the same amount
as productivity.2
Equilibrium theories of the business cycle have typically depended heavily
on intertemporal substitution of leisure to account for aggregate fluctuations
in hours worked.3 The willingness of individuals to substitute intertemporally
is measured by the elasticity of substitution between leisure in different time
periods implied by an individual’s utility function. However, the theory
developed here is able to account for large aggregate fluctuations in hours
worked relative to productivity without requiring that this elasticity be large.
This follows because the utility function of the ‘representative agent’ in our
model implies an elasticity of substitution between leisure in different periods
that is infinite.4 This result does not depend on the elasticity of substitution
implied by the preferences of the individuals who populate the economy.
Thus, the theory presented here is in principle consistent with the low estimates
of this elasticity found from studying panel data [see Altonji (1984) or MaCurdy
(1981)].
1
The fact that existing equilibrium models are inconsistent with this observation has been
stressed by Heckman (1983) and Coleman (1984).
2
Kydland and Prescott (1982) attempt to explain the above fact by including past leisure
as an argument in the individual’s utility function so as to enhance the intertemporal substitution
response to a productivity shock. However, even after introducing this feature, Kydland and
Prescott were still unable to account for this observation.
3
This is true for the technology shock theories, such as Kydland and Prescott’s (1982), as
well as the monetary shock theories of Lucas and Barro [see Lucas (1977)].
4
In this model there is a crucial distinction between the utility function of the ‘representative
agent’ and the utility function of an individual or household.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
INDIVISIBLE LABOR AND THE BUSINESS CYCLE
G.D.Hansen, Indivisible labor and the business cycle
151
311
The paper is divided as follows: The next section provides a more detailed
explanation and motivation of the indivisible labor assumption. In section 3 the
artificial economies to be studied are constructed. The first is a standard
stochastic growth model where labor is divisible, and the second introduces
indivisible labor to that economy. The second economy is a stochastic growth
version of a static general equilibrium model developed by Rogerson (1984).
Lotteries are added to the consumption set (following Rogerson) which makes
it possible to study a competitive equilibrium by solving a representative agent
problem, as in Lucas and Prescott (1971). The addition of the lotteries also
implies that the firm is providing full unemployment insurance to the workers.
The fourth section explains how the equilibrium decision rules and laws of
motion are calculated, as well as how the parameter values used when simulating
the model were chosen. Since the representative agent’s problem is not one
for which a closed form solution is available, in order to calculate decision
rules a quadratic approximation of this problem is derived using the method
described in Kydland and Prescott (1982). These equilibrium decision rules
are a set of stochastic difference equations from which the statistical properties
of the time series generated by the artificial economies can be determined.
The statistics studied are a set of standard deviations and correlations discussed
in section 5. In this section, the statistics computed using the artificial time
series are compared to the same statistics computed using U.S. time series.
Some concluding remarks are contained in section 6.
2. Motivation
Existing equilibrium theories of the business cycle analyze individuals who are
free to adjust continuously the number of hours worked (the ‘intensive margin’)
and who are always employed. There are no individuals entering or leaving
employment (the ‘extensive margin’). However, the extensive margin seems
important for explaining some aspects of labor supply at both the micro and
macro levels. Heckman and MaCurdy (1980), for example, discuss the
importance of the extensive margin for explaining female labor supply. At the
aggregate level, over half of the variation in total hours worked is due to
variation in the number of individuals employed rather than variation in average
hours worked by those employed. Consider the following decomposition of
variance involving quarterly data:
var(log Ht)=var(log ht)+var(log Nt)+2cov(log ht, log Nt),
where Ht is total hours worked, ht is average hours worked, and Nt is the
number of individuals at work, where all variables are deviations from trend.5
5
The data used for this analysis is available from the Bureau of Labor Statistics’ Labstat
data tape. The series I used were collected from households using the Current Population
Survey. For a description of the detrending method, see footnote 18.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
152
312
SOME EXTENSIONS
G.D.Hansen, Indivisible labor and the business cycle
Using this decomposition, 55% of the variance of Ht is due to variation in Nt,
while only 20% of this variance can be directly attributed to ht. The remainder
is due to the covariance term.6
Most people either work full time or not at all. This might be ascribed to
the presence of non-convexities either in individual preferences for leisure or
in the technology. For example, the technology may be such that the marginal
productivity of an individual’s work effort is increasing during the first part of
the workday or workweek, and then decreasing later on. That is, the individual
faces a production function which is convex at first and then becomes concave.
This could be due to individuals requiring a certain amount of ‘warm up’ time
before becoming fully productive. Such a technology could induce individuals
to work a lot or not at all.
Another possibility is that the non-convexity is a property of individuals’
preferences. If the utility function exhibited decreasing marginal utility of leisure
at low levels of leisure and increasing marginal utility at higher levels, individuals
would tend to choose a low level of leisure (work a lot) or use their entire time
endowment as leisure (not work at all). These preferences may be interpreted
as ‘indirect’ preferences which reflect costs associated with working each period,
such as driving a long distance to work or enduring the hassle of putting on a
suit and tie. Bearing these fixed costs makes an individual less likely to choose
to work only half a day.
In this paper the non-convexity is assumed to be a property of preferences.7
However, to make the model tractable, the non-convexity introduced—indivisible
labor—is an extreme version of the non-convexity described above. Individuals
are assumed to have preferences that are defined only at two levels of leisure—
one level corresponding to working full time and the other corresponding to
not working at all. This is modeled by assuming that the consumption
possibilities set consists of only two levels of leisure. This assumption implies
that an individual can only adjust along the extensive margin.
Of course fluctuations along both the extensive and intensive margins are
observed in the actual economy, as the above evidence indicates. However, by
studying two economies—one that exhibits fluctuations only along the intensive
margin and another with fluctuations only along the extensive margin—we can
determine the importance of non-convexities for explaining labor variability
in business cycles. If it turns out that both economies exhibit the same
cyclical behavior, then it seems likely that a model that incorporated both
margins would also exhibit similar behavior. In fact, non-convexities of this
6
Coleman (1984) comes to a similar conclusion using establishment data.
One advantage of modeling the non-convexity as a feature of the technology is that it
would likely explain why part-time workers are paid less than full-time workers, in addition
to accounting for features of the data discussed in this paper.
7
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
INDIVISIBLE LABOR AND THE BUSINESS CYCLE
G.D.Hansen, Indivisible labor and the business cycle
153
313
sort could probably be safely abstracted from when studying business cycle
phenomena. However, it happens that the two models have very different
implications and that the non-convexity improves our ability to account for
U.S. aggregate time series data.
3. Two economies
3.1. A one-sector stochastic growth model with divisible labor
The economy to be studied is populated by a continuum of identical infinitely
lived households with names on the closed interval [0, 1]. There is a single
firm with access to a technology described by a standard Cobb-Douglas
production function of the form
(1)
where labor (ht) and accumulated capital (kt) are the inputs and ␭t is a random
shock which follows a stochastic process to be described below. Agents are
assumed to observe ␭t before making any period t decisions. The assumption
of one firm is made for convenience. Since the technology displays constant
returns to scale—implying that firms make zero profit in equilibrium—the
economy would behave the same if there were many firms.
Output, which is produced by the firm and sold to the households, can
either be consumed (ct) or invested (it), so the following constraint must be
satisfied:
(2)
The law of motion for the capital stock is given by
(3)
where δ is the rate of capital depreciation. The stock of capital is owned by the
households who sell capital services to the firm.
The technology shock is assumed to follow a first-order Markov process.
In particular, ␭t obeys the following law of motion:
(4)
where the εt’s are iid with distribution function F. This distribution is assumed
to have a positive support with a finite upper bound, which guarantees that
output will always be positive. By requiring F to have mean 1-γ, the unconditional
mean of ␭t is equal to 1.
This technology shock is motivated by the fact that in post-war U.S. time
series there are changes in output (GNP) that can not be accounted for by
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
154
314
SOME EXTENSIONS
G.D.Hansen, Indivisible labor and the business cycle
changes in the inputs (capital and labor). We follow Solow (1957) and Kydland
and Prescott (1982) in interpreting this residual as reflecting shocks to
technology.
Households in this economy maximize the expected value of
where 0< β <1 is the discount factor and ct and lt are consumption and leisure
in period t, respectively. The endowment of time is normalized to be one, so
lt=1-ht. Utility in period t is given by the function
(5)
We now have a complete specification of the preferences, technology, and
stochastic structure of a simple economy where individuals are able to supply
any level of employment in the interval [0, 1]. Each period three commodities
are traded: the composite output commodity, labor, and capital services. It is
possible to consider only this sequence of spot markets since there is no
demand for intertemporal risk sharing which might exist if households were
heterogeneous.
Households solve the following problem, where wt is the wage rate at time
t and rt is the rental rate of capital:
(6)
Agents are assumed to make period t decisions based on all information available
at time t (which includes rt and wt). They have rational expectations in that
their forecasts of future wages and rental rates are the same as those implied
by the equilibrium laws of motion. The first-order conditions for the firm’s
profit maximization problem imply that the wage and rental rate each period
are equal to the marginal productivity of labor and capital, respectively.
Since there are no externalities or other distortions present in this economy,
the equal-weight Pareto optimum can be supported as a competitive equilibrium.
Since agents are homogeneous, the equal-weight Pareto optimum is the solution
to the problem of maximizing the expected welfare of the representative
agent subject to technology constraints. This problem is the following:
(7)
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
INDIVISIBLE LABOR AND THE BUSINESS CYCLE
G.D.Hansen, Indivisible labor and the business cycle
155
315
The state of the economy in period t is described by kt and ␭t. The decision
variables are ht, ct, and it.
This problem can be solved using dynamic programming techniques.8 This
requires finding the unique continuous function
(where S is the state
space) that satisfies Bellman’s equation (primes denote next period values)
(8)
where the maximization is over c and h and is subject to the same constraints
as (7). The value function, V(k, ␭), is the maximum obtainable expected
return over all feasible plans. It turns out that since the utility function is
concave and the constraint set convex, the value function is also concave.
This implies that the problem (8) is a standard finite-dimensional concave
programming problem.
Unfortunately, this problem is not one which can be solved analytically.
There is no known explicit functional form for the value function, V. In principle
this problem could be solved using numerical methods [see Bertsekas (1976)],
but a cheaper method—which does enable one to solve for closed form decision
rules—is to approximate this problem by one which consists of a quadratic
objective and linear constraints, as in Kydland and Prescott (1982). This method
will be explained briefly in section 4.
3.2. An economy with indivisible labor
The assumption of indivisible labor will now be added to the above stochastic
growth model. This will give rise to an economy where all variation in the
labor input reflects adjustment along the extensive margin. This differs
from the economy described above where all variation in the labor input
reflects adjustment along the intensive margin. In addition, the utility
function of the ‘representative agent’ for this economy will imply an elasticity
of substitution between leisure in different periods that is infinite and
independent of the elasticity implied by the utility function of the individual
households.
Indivisibility of labor is modeled by restricting the consumption possibilities
set so that individuals can either work full time, denoted by h0, or not at all.9
8
For a detailed presentation of dynamic programming methods, see Lucas, Prescott and
Stokey (1984).
9
This is consistent with the interpretation given in section 2. An alternative interpretation
of indivisible labor assumes that households can work one of two possible (non-zero) number
of hours, h1 or h2. This interpretation is consistent with an environment where each household
consists of two individuals, at least one of whom works at all times. When only one member
works, the household is working h1 hours, and when both members work the household is
working h2 hours.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
156
316
SOME EXTENSIONS
G.D.Hansen, Indivisible labor and the business cycle
In order to guarantee [using Theorem 2 of Debreu (1954)] that the solution
of the representative agent’s problem can be supported as a competitive
equilibrium, it is necessary that the consumption possibilities set be convex.
However, if one of the commodities traded is hours worked (as in the above
model), the consumption possibilities set will be non-convex. To circumvent
this problem, we convexify the consumption possibilities set by requiring
individuals to choose lotteries rather than hours worked, following Rogerson
(1984).10 Thus, each period, instead of choosing manhours, households choose
a probability of working, αt.11 A lottery then determines whether or not the
household actually works. After changing the economy in this manner, we
make it possible for the competitive equilibrium to be derived by solving a
concave programming problem, just as for the economy with divisible labor.
The new commodity being introduced is a contract between the firm and a
household that commits the household to work h0 hours with probability αt.
The contract itself is being traded, so the household gets paid whether it
works or not. Therefore, the firm is providing complete unemployment insurance
to the workers. Since all households are identical, all will choose the same
contract—that is, the same αt. However, although households are ex ante
identical, they will differ ex post depending on the outcome of the lottery: a
fraction αt of the continuum of households will work and the rest will not.12
Using (5), expected utility in period t is given by αt(log ct+A log(1-h0))+ (1αt)(log c t+A log 1). 13 This simplifies to the following function
(9)
10
In Rogerson’s paper, a static economy with indivisible labor is studied and lotteries are
introduced to solve the problem introduced by this non-convexity. Readers may wish to
consult Rogerson’s paper for a rigorous general equilibrium formulation of this type of
model.
11
Adding lotteries to the consumption set increases the choices available to households
when labor is indivisible. If lotteries were not available, households would only be able to
choose to not work (corresponding to α=0) or to work h0 (corresponding to α=1). Therefore,
adding lotteries can only make individuals better off.
12
The lottery involves drawing a realization of a random variable zt from the uniform distribution
on [0, 1]. Each individual i苸[0, 1] is now ‘renamed’ according to the following rule:
The amount worked by agent x in period t is equal to
This provides a mechanism for dividing the continuum of agents into two subsets, one where
each individual works zero hours and another where individuals work h0. The first will have
measure (1-αt) and the other measure α t. This follows from the easily verified fact that
Prob[xt(i, z)ⱕ1-αt] is equal to 1-αt for each i.
13
This uses the fact that, since preferences are separable in consumption and leisure, the
consumption level chosen in equilibrium is independent of whether the individual works or not.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
INDIVISIBLE LABOR AND THE BUSINESS CYCLE
G.D.Hansen, Indivisible labor and the business cycle
157
317
Since a fraction αt of households will work h0 and the rest will work zero, per
capita hours worked in period t is given by
(10)
The other features of this economy are exactly the same as for the economy
with divisible labor. These include the technology and the description of the
stochastic process for the technology shock. These features are described by
eqs. (1) through (4).
Firms in the economy, as in the previous economy, will want to employ
labor up to the point where fh( ␭t, kt, ht )=wt. However, due to the fact that
lottery contracts are being traded, households are not paid for the time they
actually spend working, but are instead paid for the expected amount of time
spent working. This implies that each worker is paid as if he worked ht [as
defined in (10)] rather than for the amount he actually does work. Therefore,
the budget constraint of a typical household differs from the budget constraint
for the economy where labor is divisible (6) and is given by
(11)
Thus, the problem solved by a typical household is
(12)
This problem is equivalent to the problem solved by households in a slightly
different economy where agents trade man-hours and actuarially fair insurance
contracts, rather than the type of contracts traded in the economy studied
here. In this alternative economy, which is described in more detail in the
appendix, households only get paid for the time they actually spend working.
However, if a household has purchased unemployment insurance, it will receive
compensation if the lottery determines that the household does not work. In
the appendix it is shown that households will choose to insure themselves
fully. Therefore, in equilibrium, the households will have full unemployment
insurance, just like the households populating the economy described in this
section. This implies that the equilibrium allocations for these two economies
are the same.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
158
318
SOME EXTENSIONS
G.D.Hansen, Indivisible labor and the business cycle
The following is the representative agent’s problem that must be solved to
derive the equilibrium decision rules and laws of motion:
(13)
Like problem (7), this is a standard concave discounted dynamic programming
problem. The state of the economy in period t is described by kt and ␭t. The
decision variables are αt, ct, and it.
A key property of this economy is that the elasticity of substitution between
leisure in different periods for the ‘representative agent’ is infinite. To
understand this result, first substitute ht=1-lt into (10) and solve for αt. After
substituting this expression for αt into (9) one obtains the following utility
function for the representative agent (ignoring the constant term):
(14)
where B=-A(log(l-h0))/h0. Since this utility function is linear in leisure it implies
an infinite elasticity of substitution between leisure in different periods. This
follows no matter how small this elasticity is for the individuals populating the
economy. Therefore, the elasticity of substitution between leisure in different
periods for the aggregate economy is infinite and independent of the willingness
of individuals to substitute leisure across time.14
4. Solution method and calibration
The problems (7) and (13) are not in the class of problems for which it is
possible to solve analytically for decision rules. This special class of problems
includes those with quadratic objectives and linear constraints, as well as some
other structures. For this reason, approximate economies are studied for which
the representative agent’s problem is linear-quadratic [see Kydland and Prescott
(1982)]. It is then possible to obtain explicit decision rules for these approximate
economies.
By making appropriate substitutions, one can express problems (7) and (13)
as dynamic optimization problems with decision variables it and ht and state
variables ␭t and kt. The constraints for these problems are linear although the
14
The fact that in this type of model the representative agent’s utility function is linear in
leisure was originally shown by Rogerson (1984) for his model. This result depends, however,
on the utility function being additively separable across time.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
INDIVISIBLE LABOR AND THE BUSINESS CYCLE
G.D.Hansen, Indivisible labor and the business cycle
159
319
objective functions are non-linear. For each of these problems, Kydland and
Prescott’s procedure is used to construct a quadratic approximation of the
objective function to be accurate in a neighborhood of the steady state for the
appropriate model after the technology shock has been set equal to its
unconditional mean of one.15 The reader may consult Kydland and Prescott
(1982) for details on the algorithm used for forming these approximations.16
To actually compute these quadratic approximations, solve for an
equilibrium, and generate artificial time series, it is necessary to choose a
distribution function, F, and specific parameter values for θ, δ, β, A, γ, and h0.
Kydland and Prescott (1982, 1984) follow a methodology for choosing
parameter values based on evidence from growth observations and micro
studies. This methodology will also be followed here. In fact, since they study
a similar economy, some of the above parameters (θ, δ, β) also appear in their
model. This enables me to draw on their work in selecting values for these
parameters, thereby making it easier to compare the results of the two studies.
The parameter θ corresponds to capital’s share in production. This has
been calculated using U.S. time series data by Kydland and Prescott (1982,
1984) and was found to be approximately 0.36. The rate of depreciation of
capital, δ, is set equal to 0.025 which implies an annual rate of depreciation of
10 percent. Kydland and Prescott found this to be a good compromise given
that different types of capital depreciate at different rates. The discount factor,
β, is set equal to 0.99, which implies a steady state annual real rate of interest
of four percent.
The parameter A in the utility function (5) is set equal to 2. This implies that
hours worked in the steady state for the model with divisible labor is close to
1/3. This more or less matches the observation that individuals spend 1/3 of
15
Let the steady states for the certainty version of these models be denoted by the
variable’s symbol without any subscript. Eq. (3) implies that investment in the steady state is
given by i=δk. Expressions for k and h can be determined by deriving the Euler equations for
the appropriate representative agent problem and setting ht=h, kt=k, and it=i=δk for all t. For
both economies, the steady state capital stock is given by
Hours worked in the steady state for the economy with divisible labor is given by h=(1-θ)×
(ρ+δ)/[3(ρ+δ)-θ(ρ+3δ)]; and for the economy with indivisible labor, h=(1-θ )(ρ+δ)/[␺ (ρ+δ–θ δ)]
where ␺ =-A[log(1-h0)]/h0.
16
Kydland and Prescott’s method for approximating this problem requires choosing a
vector of average deviations,
which determines the size of the neighborhood around
the steady state within which the approximation is accurate. The four components of z are
average deviations from trend of the four variables, xt=(␭t, kt, it, ht), as found in U.S. time
series data. This implies that along those dimensions where there is more variability, the
approximation will be accurate in a larger neighborhood around the steady state
. For the
exercise carried out in this paper
reflecting the average
standard deviations of these series as reported in the next section. Although attention was
paid to specifying this vector in a reasonable way, it turns out that the results are not altered
when the zt components are decreased by a factor of ten.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
160
SOME EXTENSIONS
320
G.D.Hansen, Indivisible labor and the business cycle
their time engaged in market activities and 2/3 of their time in non-market
activities.
To determine the parameter h0, I set the expressions for hours of work in
the steady state for the two models equal to each other. Since steady state
hours worked in the model with divisible labor is fully determined by the
parameters θ, δ, A, and β for which values have already been assigned (see
footnote 15), it is possible to solve for h0. This implies a value for h0 of 0.53.
The distribution function F along with the parameter γ determine the
properties of the technology shock, ␭t. The distribution of εt is assumed to be
log normal with mean (1- γ), which implies that the unconditional mean of ␭t is
1. The parameter γ is set equal to 0.95 which is consistent with the statistical
properties of the production function residual.17 The standard deviation of εt,
σε, is difficult to measure from available data since this number is significantly
affected by measurement error. A data analysis suggests that ␴e could reasonably
be expected to lie in the interval [0.007, 0.01]. A value of 0.007, for example,
would imply that a little over half of the variability in εt is being attributed to
measurement error, which is probably not unreasonable. The actual value
used for the simulations in this paper is 0.00712. This particular value was
chosen because it implies that the mean standard deviation of output for the
economy with indivisible labor is equal to the standard deviation of GNP for
the U.S. economy (see next section).
All parameters of the two models have now been determined. We are now
ready to study and compare the statistical properties of the time series generated
by these two models.
5. Results
For the purposes of this study, the statistical properties of the economies
studied are summarized by a set of standard deviations and correlations with
output that are reported in table 1.
The statistics for the U.S. economy are reported in the first two columns
of the table. Before these statistics were calculated, the time series were logged
and deviations from trend were computed. Detrending was necessary because
the models studied abstract from growth. The data were logged so that
standard deviations can be interpreted as mean percentage deviations from
17
The production function residual is measured, using U.S. time series, by
log ␭t=log yt-θ log kt-(1-θ)log h t ,
where data on GNP, capital stock (nonresidential equipment and structures), and hours
worked is obtained from a standard econometric data base. The first-order autocorrelation
coefficient for ␭t is about 0.95, indicating high serial correlation in this series. The parameter
θ was assumed to be equal to 0.36 for calculating this residual. A more detailed study of the
statistical properties of this technology shock is planned but has not yet been carried out.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
INDIVISIBLE LABOR AND THE BUSINESS CYCLE
G.D.Hansen, Indivisible labor and the business cycle
161
321
Table 1
Standard deviations in percent (a) and correlations with output (b) for U.S. and artificial
economies.
a
The U.S. time series used are real GNP, total consumption expenditures, and gross
private domestic investment (all in 1972 dollars). The capital stock series includes nonresidential
equipment and structures. The hours series includes total hours for persons at work in nonagricultural industries as derived from the Current Population Survey. Productivity is output
divided by hours. All series are seasonally adjusted, logged and detrended.
b
The standard deviations and correlations with output are sample means of statistics
computed for each of 100 simulations. Each simulation consists of 115 periods, which is the
same number of periods as the U.S. sample. The numbers in parentheses are sample standard
deviations of these statistics. Before computing any statistics each simulated time series was
logged and detrended using the same procedure used for the U.S. time series.
trend. The ‘detrending’ procedure used is the method employed by Hodrick
and Prescott (1980).18
Since much of the discussion in this section centers on the variability of
hours worked and productivity (output divided by hours worked), some
discussion of the hours series is appropriate. The time series for hours worked
used in constructing these statistics is derived from the Current Population Survey,
which is a survey of households. This series was chosen in preference to the
other available hours series which is derived from the establishment survey.
The hours series based on the household survey is more comprehensive than
18
This method involves choosing smoothed values
solve the following problem:
for the series
which
where ␭>0 is the penalty on variation, where variation is measured by the average squared
second difference. A larger value of ␭ implies that the resulting {s t} series is smoother.
Following Prescott (1983), I choose ␭=1600. Deviations from the smooth series are formed
by taking dt=xt-st.
This method is used in order to filter out low frequency fluctuations. Although other
methods (spectral techniques, for example) are available, this method was chosen because of
its simplicity and the fact that other methods lead to basically the same results [see Prescott
(1983)].
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
162
322
SOME EXTENSIONS
G.D.Hansen, Indivisible labor and the business cycle
the establishment series since self-employed workers and unpaid workers in
family-operated enterprises are included. Another advantage is that the
household series takes into account only hours actually worked rather than all
hours paid for. That is, it doesn’t include items such as paid sick leave. A
disadvantage is that the household series begins in the third quarter of 1955,
which prevented me from using data over the entire post-war period.
Sample distributions of the summary statistics describing the behavior of
the artificial economies were derived using Monte Carlo methods. The model
was simulated repeatedly to obtain many samples of artificially generated
time series. Each sample generated had the same number of periods (115) as
the U.S. time series used in the study. Before any statistics were computed,
the data were logged and the same filtering procedure applied to the U.S. data
was applied to these time series. One hundred simulations were performed
and sample statistics were calculated for each data set generated. The sample
means and standard deviations of these summary statistics are reported in the
last four columns of table 1.
When comparing the statistics describing the two artificial economies, one
discovers that the economy with indivisible labor displays significantly larger
fluctuations than the economy with divisible labor. This shows that indivisible
labor increases the volatility of the stochastic growth model for a given stochastic
process for the technology shock. In fact, it is necessary to increase ␴ε by 30
percent (from 0.00712 to 0.00929) in order to increase the standard deviation
of output for the divisible labor economy so that it is equal to the standard
deviation of GNP for the actual economy, which is 1.76. It is still the case that
0.00929 is in the interval suggested by the data (see paragraph on measuring
␴ε in the previous section). However, since it is likely that there is significant
measurement error in our empirical estimate of the production function residual,
one should prefer the lower value of ␴ε.
Another conclusion drawn from studying this table is that the fluctuations
in most variables are larger for the actual economy than for the indivisible
labor economy. It is my view that most of this additional fluctuation (except in
the case of the consumption series) is due to measurement error. Work in
progress by the author attempts to correct for measurement error in the
hours series (and hence some of the measurement error in the productivity
series).19 Preliminary findings seem to suggest that the above hypothesis is
correct. In addition, the fact that the consumption series fluctuates much more
in the actual economy than in the artificial economy can probably be explained
by the fact that nothing corresponding to consumer durables is modeled in the
economies studied here.
19
The work referred to is a chapter of my dissertation. Copies will soon be available upon
request.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
INDIVISIBLE LABOR AND THE BUSINESS CYCLE
G.D.Hansen, Indivisible labor and the business cycle
163
323
Perhaps the most significant discovery made by examining table 1 is that the
amount of variability in hours worked relative to variability in productivity is
very different for the two model economies. This relative variability can be
measured by the ratio of the standard deviation in hours worked to the standard
deviation in productivity. For the economy with indivisible labor, this ratio is
2.7, and for the economy without this feature the ratio is not significantly
above 1.20 For the U.S. economy the ratio is equal to 1.4, which is between
these two values.
As explained in the introduction, accounting for the large variability in
hours worked relative to productivity has been an open problem in equilibrium
business cycle theory. Kydland and Prescott (1982) study a version of the
stochastic growth model where labor is divisible and the utility function of
individuals is non-time-separable with respect to leisure. This non-timeseparability property is introduced to make leisure in different periods better
substitutes. However, this feature enables these authors to report a value for
this ratio of only 1.17, which is still much too low to account for the fluctuations
found in U.S. data.
On the other hand, the economy with indivisible labor studied here has
exactly the opposite problem Kyland and Prescott’s model has. The ratio
implied by this model is much larger than the ratio implied by the data.
However, this should not be surprising. In fact, it would be bothersome if this
were not the case. After all, we do observe some adjustment along the intensive
margin in the real world. Examples include workers who work overtime in
some periods and not in others or salesmen who work a different number of
hours each day. Since indivisible labor implies that all fluctuations are along
the extensive margin, one would expect—even without looking at statistics
calculated from the data—that the ratio discussed above should be somewhere
between the one implied by an indivisible labor economy and a divisible labor
economy.
6. Conclusion
A dynamic competitive equilibrium economy with indivisible labor has been
constructed with the aim of accounting for standard deviations and correlations
with output found in aggregate economic time series. Individuals in this
economy are forced to enter and exit the labor force in response to technology
shocks rather than simply adjusting the number of hours worked while
remaining continuously employed. Therefore, this is an equilibrium model
which exhibits unemployment (or employment) fluctuations in response to
aggregate shocks. Fluctuations in employment seem important for fluctuations
This ratio is still not significantly different from one even when ␴ε is increased to 0.00929.
20
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
164
324
SOME EXTENSIONS
G.D.Hansen, Indivisible labor and the business cycle
in hours worked over the business cycle since most of the variability in total
hours is unambiguously due to variation in the number employed rather than
hours per employed worker.
An important aspect of this economy is that the elasticity of substitution
between leisure in different periods for the aggregate economy is infinite and
independent of the elasticity of substitution implied by the individuals’ utility
function. This distinguishes this model, or any Rogerson (1984) style economy,
from one without indivisible labor. These include the model presented in
section 3.1 and the economy studied by Kydland and Prescott (1982). In these
divisible labor models, the elasticity of substitution for the aggregate economy
is the same as that for individuals.
This feature enables the indivisible labor economy to exhibit large
fluctuations in hours worked relative to fluctuations in productivity. Previous
equilibrium models of the business cycle, which have all assumed divisible
labor, have been unsuccessful in accounting for this feature of U.S. time series.
This is illustrated in this paper by showing that a model with divisible labor
fails to exhibit large fluctuations in hours worked relative to productivity
while the model with indivisible labor displays fluctuations in hours relative to
productivity which are much larger than what is observed. This seems to
indicate that a model which allowed for adjustment along both the extensive
margin as well as the intensive margin would have a good chance for successfully
confronting the data.
In conclusion, this study demonstrates that non-convexities such as indivisible
labor may be important for explaining the volatility of hours relative to
productivity even when individuals are relatively unwilling to substitute leisure
across time. They are also useful for increasing the size of the standard
deviations of all variables relative to the standard deviation of the technology
shock. Therefore, a smaller size shock is sufficient for explaining business
cycle fluctuations than was true for previous models such as Kydland and
Prescott’s (1982). In addition, these non-convexities make it possible for an
equilibrium model of the business cycle to exhibit fluctuations in employment.
Therefore, non-convexities will inevitably play an important role in future
equilibrium models of the cycle.
Appendix: A market for unemployment insurance
The purpose of this appendix is to show that the equilibrium of the economy
presented in section 3.2 is equivalent to the equilibrium of an economy where
labor is still indivisible but households are able to purchase any amount of
unemployment insurance they choose. In the original economy, agents are
assumed to buy and sell contracts which specify a probability of working in a
given period as opposed to buying and selling hours of work. A lottery
determines which households must work and which do not. A household is
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
INDIVISIBLE LABOR AND THE BUSINESS CYCLE
G.D.Hansen, Indivisible labor and the business cycle
165
325
paid according to the probability that it works, not according to the work it
actually does. In other words, the firm is automatically providing full
unemployment insurance to the households.
In this appendix, households choose a probability of working each period
and a lottery is held to determine which households must work, just as in the
original economy. Also, preferences, technology, and the stochastic structure
are exactly the same as for the original model. However, this economy is
different in that households only get paid for the work they actually do—
unemployed individuals get paid nothing by the firm. But, the household does
have access to an insurance market which preserves the complete markets
aspect of the original model. It is shown below that the equilibrium of this
economy is equivalent to that of the original economy since individuals will
choose to be fully insured in equilibrium. This is shown by proving that the
problem solved by households is the same as the problem solved by households
(12) in the original model.
The problem solved by the households can be described as follows: Each
period, households choose a probability of working, αt, a level of unemployment
compensation, yt, and consumption and investment contingent on whether the
household works or not, cst and ist (s=1, 2). These are chosen to solve the
following dynamic programming problem (primes denote next period values):
(A.1)
subject to
(A.2)
(A.3)
(A.4)
The function V(␭, K, k) is the value function which depends on the
household’s state. The state vector includes the capital owned by the household,
plus the economy wide state variables ␭ and K, where K is the per capita capital
stock.21 The functions w(␭, K) and r(␭, K) are the wage rate and rental rate
21
Since we are allowing households to choose any level of unemployment insurance they
wish, we have to allow for the heterogeneity that may come about because different households
will have different income streams. This is why the distinction is made between the per capita
capital stock, K, and the households accumulated capital stock, k. However, this heterogeneity
will disappear in equilibrium since all households will choose full insurance, so K=k in
equilibrium.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
166
326
SOME EXTENSIONS
G.D.Hansen, Indivisible labor and the business cycle
of capital respectively, and p(α) is the price of insurance, which is a function of
the probability that the household works. Also, since individuals’ preferences
are the same as for the original model, u(c)=log c and v(l)=A log l.
The insurance company in this economy maximizes expected profits which
are given by p(α)y-(1-α)y. That is, the firm collects revenue p(α)y and pays y
with probability 1-α. To guarantee that profits are bounded, p(α)=(1-α).
Therefore, the price the household must pay for insurance equals the probability
that the household will collect on the insurance.
One can now substitute this expression for p into constraints (A.2) and
(A.3). After eliminating the constraints by substituting out is and cs (s=1, 2), one
can write the following first-order necessary conditions for and y:
(A.5)
(A.6)
Eq. (A.6) implies, given the strict concavity of u, that c 1=c 2 . This plus eq.
(A.5) imply that
This, in turn, implies that i1=i2. Therefore, the lefthand sides of eqs. (A.2) and (A.3) are identical. Since these constraints will be
binding in equilibrium, y will be chosen so that the right-hand sides are equal
as well. This means that y=wh0 in equilibrium. That is, households will choose
to insure themselves fully. This has the implication that all households will
choose the same sequence of capital stocks, so K=k.
Substituting these results into the household’s optimization problem (A.1)
yields the following problem: Households choose c, i, k′ , and α to
(A.7)
This problem is identical to problem (12). Therefore, the equilibrium
allocation for the original economy, where the firm provides full unemployment
insurance to workers by assumption, is equivalent to the equilibrium allocation
for an economy where households get paid by the firm only for work done
but have access to a risk-neutral insurance market. This result, of course,
depends crucially on the probability α being publicly observable and the contract
being enforceable. That is, it must be the case that the agent announces the
same α to both the firm and the insurance company, and if the agent loses the
lottery (that is, has to work) this is known by all parties. For example, this result
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
INDIVISIBLE LABOR AND THE BUSINESS CYCLE
G.D.Hansen, Indivisible labor and the business cycle
167
327
would not hold if α depended on some underlying choice variable like effort
that was not directly observed by the insurance company. In this case a difficult
moral hazard problem would arise.
References
Altonji, J.G., 1984, Intertemporal substitution in labor supply: Evidence from micro data,
Unpublished manuscript (Columbia University, New York).
Ashenfelter, O., 1984, Macroeconomic analyses and microeconomic analyses of labor supply,
Carnegie-Rochester Conference Series on Public Policy 21, 117–156.
Bertsekas, D.P., 1976, Dynamic programming and stochastic control (Academic Press, New
York).
Coleman, T.S., 1984, Essays on aggregate labor market business cycle fluctuations,
Unpublished manuscript (University of Chicago, Chicago, IL).
Debreu, G., 1954, Valuation equilibrium and Pareto optimum, Proceedings of the National
Academy of Sciences 40, 588–592.
Heckman, J.J., 1984, Comments on the Ashenfelter and Kydland papers, Carnegie-Rochester
Conference Series on Public Policy 21, 209–224.
Heckman, J.J. and T.E.MaCurdy, 1980, A life cycle model of female labor supply, Review of
Economic Studies 47, 47–74.
Hodrick, R.J. and E.C.Prescott, 1980, Post-war U.S. business cycles: An empirical investigation.
Working paper (Carnegie-Mellon University, Pittsburgh, PA).
Kydland, F.E. and E.C.Prescott, 1982, Time to build and aggregate fluctuations, Econometrica
50, 1345–1370.
Kydland, F.E. and E.C.Prescott, 1984, The workweek of capital and labor. Unpublished
manuscript (Federal Reserve Bank of Minneapolis, Minneapolis, MN).
Lucas, R.E., Jr., 1977, Understanding business cycles, Carnegie-Rochester Conference Series
on Public Policy 5, 7–29.
Lucas, R.E., Jr. and E.C.Prescott, 1971, Investment under uncertainty, Econometrica 39,
659–681.
Lucas, R.E., Jr., E.C.Prescott and N.L.Stokey, 1984, Recursive methods for economic
dynamics, Unpublished manuscript (University of Minnesota, Minneapolis, MN).
MaCurdy, T.E., An empirical model of labor supply in a life-cycle setting, Journal of Political
Economy 89, 1059–1085.
Rogerson, R., Indivisible labour, lotteries and equilibrium, Unpublished manuscript
(University of Rochester, Rochester, NY).
Prescott, E.C., Can the cycle be reconciled with a consistent theory of expectations? or a
progress report on business cycle theory, Unpublished manuscript (Federal Reserve
Bank of Minneapolis, Minneapolis, MN).
Solow, R.M., Technical change and the aggregate production function, The Review of
Economics and Statistics 39, 312–320.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
168
CHAPTER 9
Federal Reserve Bank of Minneapolis
Quarterly Review Spring 1992
The Labor Market in Real Business Cycle Theory*
Gary D.Hansen
Professor of
Economics
University of California,
Los Angeles
Randall Wright
Senior Economist
Research Department
Federal Reserve Bank of Minneapolis
and Professor of Economics
University of Pennsylvania
The basic objective of the real business cycle research
program is to use the neoclassical growth model to
interpret observed patterns of fluctuations in overall
economic activity. If we take a simple version of the
model, calibrate it to be consistent with long-run
growth facts, and subject it to random technology
shocks calibrated to observed Solow residuals, the
model displays short-run cyclical behavior that is
qualitatively and quantitatively similar to that
displayed by actual economies along many important
dimensions. For example, the model predicts that
consumption will be less than half as volatile as output,
that investment will be about three times as volatile as
output, and that consumption, investment, and
employment will be strongly positively correlated with
output, just as in the postwar U.S. time series.1 In this
sense, the real business cycle approach can be thought of
as providing a benchmark for the study of aggregate
fluctuations.
In this paper, we analyze the implications of real
business cycle theory for the labor market. In
particular, we focus on two facts about U.S. time series:
the fact that hours worked fluctuate considerably more
than productivity and the fact that the correlation
between hours worked and productivity is close to zero.2
These facts and the failure of simple real business cycle
models to account for them have received considerable
attention in the literature. [See, for example, the
extended discussion by Christiano and Eichenbaum
(1992) and the references they provide.] Here we first
document the facts. We then present a baseline real
business cycle model (essentially, the divisible labor
The Editorial Board for this paper was V.V.Chari, Preston
J.Miller, Richard Rogerson, and Kathleen S.Rolfe.
model in Hansen 1985) and compare its predictions
with the facts. We then consider four extensions of the
baseline model that are meant to capture features of the
world from which this model abstracts. Each of these
extensions has been discussed in the literature.
However, we analyze them in a unified framework with
common functional forms, parameter values, and so on,
so that they can be more easily compared and evaluated
in terms of how they affect the model’s ability to explain
the facts.
The standard real business cycle model relies
exclusively on a single technology shock to generate
fluctuations, so the fact that hours worked vary more
than productivity implies that the short-run labor
supply elasticity must be large. The first extension of
the model we consider is to recognize that utility may
depend not only on leisure today but also on past
leisure; this possibility leads us to introduce nonseparable
preferences, as in Kydland and Prescott 1982. 3 This
extension of the baseline model has the effect of
increasing the relevant elasticity, by making households
more willing to substitute leisure in one period for
leisure in another period in response to short-run
productivity changes. At the same time, with these
* This paper is also available in Spanish in Cuadernos Economicos de
ICE, a quarterly publication of the Ministerio de Economía y Hacienda.
The paper appears here with the permission of that publication’s editor,
Manuel Santos.
1
These properties are also observed in other countries and time periods.
See Kydland and Prescott 1990 for an extended discussion of the postwar
U.S. data, and see Blackburn and Ravn 1991 or Backus and Kehoe,
forthcoming, for descriptions of other countries and time periods.
2
Although we concentrate mainly on these cyclical facts, we also
mention an important long-run growth fact that is relevant for much of our
discussion: total hours worked per capita do not display trend growth
despite large secular increases in average productivity and real wages.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
LABOR MARKETS AND THE CYCLE
169
Gary D.Hansen, Randall Wright
Real Business Cycle Theory
preferences, households do not increase their work
hours in response to permanent productivity growth.
Thus, the nonseparable leisure model generates an
increased standard deviation of hours worked relative
to productivity without violating the long-run growth
fact that hours worked per capita have not increased
over long periods despite large increases in
productivity.
The second extension of the baseline real business
cycle model we consider is to assume that labor is
indivisible, so that workers can work either a fixed
number of hours or not at all, as in Hansen 1985. In this
version of the model, all variation in the labor input
must come about by changes in the number of employed
workers, which is the opposite of the standard model,
where all variation comes about by changes in hours
per worker. Although the data display variation along
both margins, the indivisible labor model is perhaps a
better abstraction, since the majority of the variance in
the labor input in the United States can be attributed to
changes in the number of employed workers. In the
equilibrium of the indivisible labor model, individual
workers are allocated to jobs randomly, and this turns
out to imply that the aggregate economy displays a
large labor supply elasticity even though individual
hours do not respond at all to productivity or wage
changes for continuously employed workers. The large
aggregate labor supply elasticity leads to an increased
standard deviation of hours relative to productivity, as
compared to the baseline model.
Neither nonseparable utility nor indivisible labor
changes the result that the real business cycle model
implies a large positive correlation between hours and
productivity while the data display a near-zero
correlation. This result arises because the model is
driven by a single shock to the aggregate production
function, which can be interpreted as shifting the labor
demand curve along a stable labor supply curve and
inducing a very tight positive relationship between
hours and productivity. Hence, the next extension we
consider is to introduce government spending shocks, as in
Christiano and Eichenbaum 1992. If public
consumption is an imperfect substitute for private
consumption, then an increase in government spending
has a negative wealth effect on individuals, which
induces them to work more if leisure is a normal good.
Therefore, government spending shocks can be
interpreted as shifting the labor supply curve along the
labor demand curve. Depending on the size of and the
response to the two shocks, with this extension the
model can generate a pattern of hours versus
productivity closer to that found in the data.
The final extension we consider is to introduce
household production as in Benhabib, Rogerson, and
Wright 1991. The basic idea is to recognize that agents
derive utility from home-produced as well as marketproduced consumption goods and derive disutility from
working in the home as well as in the market. In this
version of the model, individuals, by working less at
home, can increase hours of market work without
reducing leisure as much. Therefore, the addition of
household production increases the short-run labor
supply elasticity and the standard deviation of hours
relative to productivity. Furthermore, to the extent that
shocks to household production are less than perfectly
correlated with shocks to market production,
individuals will have an incentive to substitute between
home and market activity at a point in time. This is in
addition to the standard incentive to substitute between
market activity at different dates. Therefore, home
production shocks, like government spending shocks,
shift the labor supply curve and can generate a pattern
of hours versus productivity closer to that found in the
data.
Our basic finding is that each of these four
extensions to the baseline real business cycle model
improves its performance quantitatively, even though
the extensions work through very different economic
channels. As will be seen, some of the resulting models
seem to do better than others along certain dimensions,
and some depend more sensitively than others on
parameter values. Our goal here is not to suggest that
one of these models is best for all purposes; which is
best for any particular application will depend on the
context. Rather, we simply want to illustrate here how
incorporating certain natural features into the standard
real business cycle model affects its ability to capture
some key aspects of labor market behavior.
The Facts
In this section, we document the relevant business cycle
facts. We consider several measures of hours worked
and productivity and two sample periods (since some of
the measures are available only for a shorter period).
As in Prescott 1986, we define the business cycle as fluctuations around some slowly moving trend. For any given
data series, we first take logarithms and then use the
Hodrick-Prescott filter (as described in Prescott 1986)
to remove the trend.
Table 1 contains some summary statistics for
quarterly U.S. data that are computed from deviations
constructed in this manner. The sample period is from
1955:3 to 1988:2. The variables are y=output,
c=consumption (nondurables plus services), i=fixed
investment, h=hours worked, and w =average
productivity (output divided by hours worked).4 For
3
Note that these preferences are nonseparable between leisure in
different periods; they may or may not be separable between leisure and
consumption in a given period.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
170
SOME EXTENSIONS
Tables 1 and 2
Cyclical Properties of U.S. Time Series
Table 2 1947:1–1991:3
* All series are quarterly, are in 1982 dollars, and have been logged and detrended with the Hodrick-Prescott filter. The output series, y, is the gross national product; c is
consumption of nondurables and services; and i is fixed investment. Productivity is w=y/h.
Sources: Citicorp’s Citibase data bank and Hansen 1991
4
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
LABOR MARKETS AND THE CYCLE
171
Gary D.Hansen, Randall Wright
Real Business Cycle Theory
each variable j, we report the following statistics: the
(percent) standard deviation σj, the standard deviation
relative to that of output σj /σy, and the correlation with
output cor(j, y). We also report the relative standard
deviation of hours to that of productivity σh/σw and the
correlation between hours and productivity cor(h, w).
We present statistics for four measures of h and w.
Hours series 1 is total hours worked as recorded in the
household survey and covers all industries. Hours series
2 is total hours worked as recorded in the establishment
survey and covers only nonagricultural industries. These
two hours series could differ for two reasons: they are
from different sources, and they cover different
industries.5 To facilitate comparison, we also report, in
hours series 3, hours worked as recorded in the household
survey but only for nonagricultural industries. Finally,
hours series 4 is a measure of hours worked in efficiency
units.6
The reason for the choice of 1955:3–1988:2 as the
sample period is that hours series 3 and 4 are only
available for this period. However, the other series are
available for 1947:1–1991:3, and Table 2 reports
statistics from this longer period for the available
variables.
Both Table 1 and Table 2 display the standard
business cycle facts. All variables are positively
correlated with output. Output is more variable than
consumption and less variable than investment. Hours
are slightly less variable than or about as variable as
output, with σh/σy ranging between 0.78 and 1.01,
depending on the hours series and the period. Overall,
all variables are more volatile in the longer period, but
the relative volatilities of the variables are about the
same in the two periods. (An exception is investment,
which looks somewhat less volatile relative to output in
the longer period.)
We want to emphasize two things. First, hours
fluctuate more than productivity, with the magnitude of
σh/σw ranging between 1.37 and 2.15, depending on the
series and the period. Second, the correlation between
hours and productivity is near zero or slightly negative,
with cor(h, w) ranging between -0.35 and 0.10, depending
on the series and the period. Chart 1 shows the scatter
plot of h versus w from hours series 1 for the longer
sample period (Plots from the other hours series look
similar.)
The Standard Model
In this section, we present a standard real business cycle
model and investigate its implications for the facts just
described.
The model has a large number of homogeneous
households. The representative household has preferences
defined over stochastic sequences of consumption ct and
leisure lt, described by the utility function
(1)
where E denotes the expectation and β the discount
factor, with β苸(0,1). The household has one unit of
time each period to divide between leisure and hours of
work:
(2)
The model has a representative firm with a constant
returns-to-scale Cobb-Douglas production function that
uses capital kt and labor hours ht to produce output yt:
(3)
where θ is the capital share parameter and
is a
stochastic term representing random technological
progress. In general, we would assume that
where is a constant yielding exogenous deterministic
growth and zt evolves according to the process
(4)
where ρ苸(0,1) and εt is independent and normally distributed with mean zero and standard deviation σε.
However, in this paper, we abstract from exogenous
growth by setting
0.7 Capital evolves according to
the law of motion
(5)
where δ is the depreciation rate and it investment. Finally, the economy must satisfy the resource constraint
(6)
We are interested in the competitive equilibrium of
this economy. Since externalities or other distortions
are not part of this model (or the other models that we
consider), the competitive equilibrium is efficient. Hence,
we can determine the equilibrium allocation by solving
the social planner’s problem of maximizing the repre4
We use the letter w because average productivity is proportional to
marginal productivity (given our functional forms), which equals the real
wage rate in our models.
5
The establishment series is derived from payroll data and measures
hours paid for, while the household series is taken from a survey of workers
that attempts to measure hours actually worked. These two measures
could differ, for example, because some workers may be on sick leave or
vacation but still get paid. The household series is a better measure of the
labor input, in principle, but because it is based on a survey of workers
rather than payroll records, it is probably less accurate.
6
Efficiency units are constructed from hours series 3 by disaggregating
individuals into age and sex groups and weighting the hours of each group
by its relative hourly earnings; see Hansen 1991 for details.
7
Adding exogenous growth does not affect any of the statistics we
report (as long as the parameters are recalibrated appropriately) given the
way we filter the data; therefore, we set
in order to simplify the
presentation. See Hansen 1989.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
172
SOME EXTENSIONS
sentative agent’s expected utility subject to feasibility
constraints. That problem in this case is to maximize U
subject to equations (2)–(6) and some initial conditions
(k0, z0). The solution can be represented as a pair of
stationary decision rules for hours and investment,
ht=h*(kt, zt) and it=i*(kt, zt), that determine these two
variables as functions of the current capital stock and
technology shock. The other variables, such as consumption and output, can be determined from the decision rules using the constraints, while prices can be
determined from the relevant marginal conditions.
Standard numerical techniques are used to analyze
the model. We choose functional forms and parameter
values and substitute the constraint ct+it=f(zt, kt, ht) into
the instantaneous return function u to reduce the problem
to one of maximizing an objective function subject to
linear constraints. Then we approximate the return
function with a quadratic return function by taking a
Taylor’s series expansion around the deterministic steady
state. The resulting linear-quadratic problem can be
easily solved for optimal linear decision rules, ht= h(kt,
zt) and it=i(kt, zt); see Hansen and Prescott 1991 for
details. Using these decision rules, we simulate the
model, take logarithms of the artificially generated data,
apply the Hodrick-Prescott filter, and compute statistics
on the deviations (exactly as we did to the actual time
series). We run 100 simulations of 179 periods (the
number of quarters in our longer data set) and report
the means of the statistics across these simulations.
Preferences are specified so that the model is able to
capture the long-run growth fact that per-capita hours
worked display no trend despite large increases in
productivity and real wages. When preferences are time
separable, capturing this fact requires that the
instantaneous utility function satisfy
(7)
or
(8)
where σ is a nonzero parameter and ␯(l) is an increasing
and concave function. (See King, Plosser, and Rebelo
1987, for example.) Intuitively, the growth facts imply
that the wealth and substitution effects of long-run
changes in productivity cancel, so the net effect is that
hours worked do not change.8 We consider only preferences that satisfy (7) or (8); in fact, for convenience, we
assume that
(9)
Parameter values are calibrated as follows. The
discount factor is set to β=0.99 so as to imply a
reasonable steady-state real interest rate of 1 percent
per period (where a period is one quarter). The capital
share parameter is set to θ=0.36 to match the average
fraction of total income going to capital in the U.S.
economy. The depreciation rate is set to δ=0.025, which
(given the above-mentioned values for β and θ) implies
a realistic steady-state ratio of capital to output of
about 10 and a ratio of investment to output of 0.26.
The parameter A in the utility function (9) is chosen so
that the steady-state level of hours worked is exactly
h=1/3, which matches the fraction of discretionary time
spent in market work found in time-use studies (for
example, Juster and Stafford 1991). Finally, the parameter
ρ in (4) is set to ρ=0.95, and the standard deviation of
ε is set to σε=0.007, which are approximately the values
settled on by Prescott (1986).
We focus on the following statistics generated by
our artificial economy: the standard deviation of output;
the standard deviations of consumption, investment,
and hours relative to the standard deviation of output;
the ratio of the standard deviation of hours to the
standard deviation of productivity; and the correlation
between hours and productivity. The results are shown
in Table 3, along with the values for the same statistics
from our longer sample from the U.S. economy (from
Table 2). We emphasize the following discrepancies
between the simulated and actual data. First, the model
has a predicted standard deviation of output which is
considerably less than the same statistic for the U.S.
economy in either period. Second, the model predicts
that σh/σw is less than one, while it is greater than one in
the data. Third, the correlation between hours and
productivity in the model is far too high.
The result that output is not as volatile in the
model economy as in the actual economy is not too
surprising, since the model relies exclusively on a
single technology shock, while the actual economy is
likely to be subject to other sources of uncertainty as
well. The result that in the model hours worked do not
fluctuate enough relative to productivity reflects the
fact that agents in the model are simply not sufficiently
willing to substitute leisure in one period for leisure in
8
Other specifications can generate a greater short-run response of hours
worked to productivity shocks; but while this is desirable from the point of
view of explaining cyclical observations, it is inconsistent with the growth
facts. For example, the utility function used in Greenwood, Hercowitz, and
Huffman 1988, u(c, l)=␯(c+Al), has a zero wealth effect and hence a large
labor supply elasticity, but implies that hours worked increase over time with
productivity growth. This specification is consistent with balanced growth
if we assume the parameter A grows at the same average rate as technology.
Although such an assumption may seem contrived, it can be justified as the
reduced form of a model with home production in which the home and
market technologies advance at the same rate on average, as shown in
Greenwood, Rogerson, and Wright 1992.
6
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
LABOR MARKETS AND THE CYCLE
173
Gary D.Hansen, Randall Wright
Real Business Cycle Theory
Table 3
Cyclical Properties of U.S. and Model-Generated Time Series
* U.S. data here are the same as those in Table 2; they are for the longer time period: 1947:1–1991:3.
** The standard deviations and correlations computed from the models’ artificial data are the sample means of statistics computed for each of 100 simulations.
Each simulation has 179 periods, the number of quarters in the U.S. data.
Source: Citicorp’s Citibase data bank
other periods. Finally, the result that hours and
productivity are too highly correlated in the model reflects
the fact that the only impulse driving the system is the
aggregate technology shock.
Chart 2 depicts the scatter plot between h and w
generated by the model. Heuristically, Chart 2 displays
a stable labor supply curve traced out by a labor demand
curve shifting over time in response to technology shocks.
This picture obviously differs from that in Chart 1.
Nonseparable Leisure
Following Kydland and Prescott (1982), we now attempt to incorporate the idea that instantaneous utility
might depend not just on current leisure, but rather on
a weighted average of current and past leisure. Hotz,
Kydland, and Sedlacek (1988) find evidence in the panel
data that this idea is empirically plausible. One interpretation they discuss concerns the fact that individuals need to spend time doing household chores, making
repairs, and so on, but after doing so they can neglect
these things for a while and spend more time working
in the market until the results of their home work
depreciate. The important impact of a nonseparable
utility specification for our purposes is that, if leisure
in one period is a relatively good substitute for leisure
in nearby periods, then agents will be more willing to
substitute intertemporally, and this increases the shortrun labor supply elasticity.
Assume that the instantaneous utility function is
u(ct, Lt)=log(ct)+Alog(Lt), where Lt is given by
(10)
and impose the restriction that the coefficients ai sum to
one. If we also impose the restriction that
(11)
for i=1, 2, …, so that the contribution of past leisure to Lt decays geometrically at rate η, then the
two parameters a 0 and η determine all of the coefficients in (10). Since Lt, and not simply lt, now provides utility, individuals are more willing to
intertemporally substitute by working more in some
periods and less in others. (At the same time, in a
deterministic steady state or along a deterministic
7
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
174
SOME EXTENSIONS
Charts 1–5
Hours Worked vs. Productivity in the Data and the Models
8
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
LABOR MARKETS AND THE CYCLE
175
Gary D.Hansen, Randall Wright
Real Business Cycle Theory
balanced growth path, this model delivers the correct
prediction concerning the effect of productivity growth
on hours worked.)
The equilibrium can again be found as the solution
to a social planner’s problem, which in this case
maximizes U subject to (2)–(26), (10)–(11), and initial
conditions. 9 The parameter values we use for the
preference structure are a0=0.35 and η=0.10, which are
the values implied by the estimates in Hotz, Kydland,
and Sedlacek 1988; other parameter values are the same
as in the preceding section.
The results are in Table 3. Notice that output is
more volatile here than in the standard model, with σy
increasing from 1.30 to 1.51. Also, the standard
deviation of hours worked relative to that of productivity
has increased considerably, to σh/σw=1.63, and the
correlation between hours and productivity has decreased
somewhat to 0.80. Chart 3 depicts the scatter plot of h
versus w generated by this model. Although these points
trace out a labor supply curve that is flatter than the one
in Chart 2, the model still does not generate the cloud in
Chart 1. We conclude that introducing nonseparable
leisure improves things in terms of σh/σw, but does little
for cor(h, w).
Indivisible Labor
We now take up the indivisible labor model of Hansen
(1985), in which individuals are constrained to work
either zero or hˆ hours in each period, where 0<hˆ<1.
Adding this constraint is meant to capture the idea that
the production process has important nonconvexities or
fixed costs that may make varying the number of employed workers more efficient than varying hours per
worker. As originally shown by Rogerson (1984, 1988),
in the equilibrium of this model, individuals will be
randomly assigned to employment or unemployment
each period, with consumption insurance against the
possibility of unemployment. Thus, this model generates fluctuations in the number of employed workers
over the cycle. As we shall see, it also has the feature
that the elasticity of total hours worked increases relative to the standard model.
Let πt be the probability that a given agent is
ˆ
employed in period t, so that Ht=πth is per-capita hours
worked if we assume a large number of ex ante identical
agents. Also, let c0t denote the consumption of an
unemployed agent and c1t the consumption of an employed
agent. As part of the dynamic social planning problem,
πt, c0t, and c1t are chosen to maximize in each period, to
(12)
the following constraint:
(13)
where ct is total per-capita consumption. When u(c,
l)=log(c) +Alog(l), the solution can be shown to imply
that c1t=c0t= ct.10
Therefore, in the case under consideration, expected
utility can be written
(14)
where B≡-Alog(1-hˆ)/hˆ>0 and, as defined above, Ht is
hours worked per capita. Therefore, the indivisible labor model is equivalent to a divisible labor model with
preferences described by
(15)
where
Based on this equiva-
lence, we can solve the indivisible labor model as if it
were a divisible labor model with a different instantaneous utility function, by maximizing subject to (2)–
(6) and initial conditions.11
Two features of the indivisible labor economy bear
mention. First, as discussed earlier, fluctuations in the
labor input come about by fluctuations in employment
rather than fluctuations in hours per employed worker.
This is the opposite of the standard model and is perhaps
preferable, since the majority of the variance in total
hours worked in the U.S. data is accounted for by
variance in the number of workers. 12 Second, the
indivisible labor model generates a large intertemporal
substitution effect for the representative agent because
instantaneous utility,
is linear in H, and therefore
the indifference curves between leisure in any two periods
are linear. This is true despite the fact that hours worked
are constant for a continuously employed worker.
Return to Table 3 for the results of our simulations
of this model. 13 The indivisible labor model is
considerably more volatile than the standard model,
with σy increasing from 1.30 to 1.73. Also, σh/σw has
9
For the solution techniques that we use, this problem is expressed as
a dynamic program. The stock of accumulated past leisure is defined to be
Xt, and we write
Lt=a0lt+η(1-a0)X t
Xt+1=(1-η)X t+l t.
These equations replace (10) and (11) in the recursive formulation.
10
This implication follows from the fact that u is separable in c and l and
does not hold for general utility functions; see Rogerson and Wright 1988.
11
Since the solution to the planner’s problem in the indivisible labor
model involves random employment, we need to use some type of lottery
or sunspot equilibrium concept to support it as a decentralized equilibrium;
see Shell and Wright, forthcoming.
12
See Hansen 1985 for the U.S. data. Note, however, that European
data display greater variance in hours per worker than in the number of
workers; see Wright 1991, p. 17.
13
The new parameter B is calibrated so that steady-state hours are
again equal to 1/3; the other parameters are the same as in the standard
model.
9
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
176
SOME EXTENSIONS
increased from 0.94 to 2.63, actually somewhat high
when compared to the U.S. data. Of course, this model
is extreme in the sense that all fluctuations in the labor
input result from changes in the number of employed
workers, and models in which both the number of
employed workers and the number of hours per worker
vary fall somewhere between the standard divisible labor
model and the indivisible labor model with respect to
this statistic. (See Kydland and Prescott 1991 or Cho
and Cooley 1989, for example.) Finally, the model
implies that cor(h, w)=0.76, slightly lower than the
models discussed above but still too high. For the sake
of brevity, the scatter plot between h and w is omitted;
for the record, it looks similar to the one in Chart 3,
although the indivisible labor model displays a little
more variation in hours worked.
Government Spending
We now introduce stochastic government spending, as
in Christiano and Eichenbaum 1992. (That paper also
provides motivation and references to related work.)
Assume that government spending, gt, is governed by
(16)
respectively. (In addition, the average of gt/yt in our
sample, which is 0.22, is used to calibrate )
For the results, turn again to Table 3. The
government spending model actually behaves very much
like the standard model, except that the correlation
between hours and productivity decreases to cor(h,
w)=0.49, which is better than the previous models
although still somewhat larger than the U.S. data. Chart
4 displays the scatter plot generated by the model with
only government spending shocks (that is, with the
variance in the technology shock set to σε=0), and Chart
5 displays the scatter plot for the model with both
shocks. These charts illustrate the intuition behind the
results: technology shocks shift labor demand and trace
out the labor supply curve, government shocks shift
labor supply and trace out the labor demand curve, and
both shocks together generate a combination of these
two effects. The net results will be somewhat sensitive
to the size of and the response to the two shocks; however,
for the estimated parameter values, this model generates
a scatter plot that is closer to the data than does the
standard model.15
Home Production
where ␭⑀(0, 1) and µt is independent and normally distributed with mean zero and standard deviation σµ.
Furthermore, as in Christiano and Eichenbaum 1992,
assume that µt is independent of the technology shock.
Also assume that government spending is financed by
lump-sum taxation and that it enters neither the utility
function nor the production function.14 Then the equilibrium allocation for the model can be found by solving
the planner’s problem of maximizing U subject to (16),
(2)–(5), and, instead of (6), the new resource constraint
(17)
An increase in gt is a pure drain on output here. Since
leisure is a normal good, the negative wealth effect of
an increase in gt induces households to work more.
Intuitively, shocks to gt shift the labor supply curve
along the demand curve at the same time that technology shocks shift the labor demand curve along the
supply curve. This first effect produces a negative relationship between hours and productivity, while the
second effect produces a positive relationship. The net
effect on the correlation between hours and productivity in the model depends on the size of the gt shocks
and on the implied wealth effect, which depends,
among other things, on the parameter λ in the law of
motion for g t (because temporary shocks have a
smaller wealth effect than permanent shocks). Hence,
the calibration of this law of motion is critical. An
ordinary least squares regression based on equation
(16) yields estimates for λ and σµ of 0.96 and 0.021,
We now consider the household production model analyzed in Benhabib, Rogerson, and Wright 1991. (That
paper also provides motivation and references to related work.)
Instantaneous utility is still written u(c,
l)=log(c)+Alog(l), but now consumption and leisure have
a different interpretation. We assume that
(18)
(19)
14
A generalization is to assume that instantaneous utility can be
written u(C, l), where C=C(c, g) depends on private consumption and
government spending. The special case where C=c is the one we consider
here, while the case where C=c+g can be interpreted as the standard model,
since then increases in g can be exactly offset by reductions in c and the other
variables will not change. Therefore, the model with C= c+g generates
exactly the same values of all variables, except that c+g replaces c. The
assumption that c and g are perfect substitutes implies that they are perfectly
negatively correlated, however. A potentially interesting generalization
would be to assume that
C(c, g)=[αcφ+(1-α)gφ]1/φ
where l/(l-φ) is the elasticity of substitution.
15
The size of the wealth effect depends on the extent to which public
consumption and private consumption are substitutes. For example, if they
were perfect substitutes, then a unit increase in g would simply crowd out
a unit of c with no effect on hours worked or any of the other endogenous
variables. We follow Christiano and Eichenbaum 1992 in considering the
extreme case where g does not enter utility at all. Also, the results depend
on the (counterfactual) assumption that the shocks to government spending
and technology are statistically independent. Finally, the results depend on
the estimates of the parameters in the law of motion (16). The estimates in
the text are from the period 1947:1–1991:3 and are close to the values used
in Christiano and Eichenbaum 1992. Estimates from our shorter sample
period, 1955:3–1988:2, imply a higher λ of 0.98 and a lower σµ of 0.012,
which in simulations yield cor(h, w)=0.65.
10
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
LABOR MARKETS AND THE CYCLE
Gary D.Hansen, Randall Wright
Real Business Cycle Theory
where cMt is consumption of a market-produced good, cHt
is consumption of a home-produced good, hMt is hours
worked in the market sector, and hHt is hours worked in
the home, all in period t. Notice that the two types of
work are assumed to be perfect substitutes, while the
two consumption goods are combined by an aggregator
that implies a constant elasticity of substitution equal to
1/(1-e).
This model has two technologies, one for market
production and one for home production:
(20)
(21)
where ␪ and ␩ are the capital share parameters. The
two technology shocks follow the processes
(22)
177
σh/σw=1.92. And cor(h, w) has decreased to 0.49, the
same as in the model with government spending.17
The intuition behind these results is that agents
substitute in and out of market activity more in the
home production model than in the standard model
because they can use nonmarket activity as a buffer. The
degree to which agents do this depends on their
willingness to substitute cM for cH, as measured by e, and
on their incentive to move production between the two
sectors, as measured by γ. (Lower values of γ entail
more frequent divergence between zM and zH and, hence,
more frequent opportunities to specialize over time.)
Note that some aspects of the results do not actually
depend on home production being stochastic.18 However,
the correlation between productivity and market hours
does depend critically on the size of the home technology
shock, exactly as it depends on the size of the second
shock in the government spending model. We omit the
home production model’s scatter plot between h and w,
but it looks similar to that of the model with government
shocks.
Conclusion
(23)
where the two innovations are normally distributed
with standard deviations σM and σH, have a contemporaneous correlation γ=cor(εMt, εHt), and are independent
over time. In each period, a capital constraint holds:
kMt+kHt=kt, where total capital evolves according to
kt+1=(1-δ)kt+it. Finally, the constraints imply that all
(24)
(25)
new capital is produced in the market sector.
The parameters β, ␪, ␦, and ␳ are set to the
values used in the previous sections. The two utility
parameters A and a are set to deliver steady-state
values of hM=0.33 and hH=0.28, as found in the timeuse studies (Juster and Stafford 1991), and the capital
share parameter in the household sector is set to
␩=0.08, implying a steady-state ratio of c H/c M of
approximately 1/4.16 The variances of the two shocks
are assumed to be the same: ␴H=␴M =0.007. The
parameter e, which determines the elasticity of
substitution between cM and cH, and γ, which is the
correlation between ε M and ε H , are set to the
benchmark values used in Benhabib, Rogerson, and
Wright 1991: e=0.8 and γ=2/3.
The results are at the bottom of Table 3. In the
home production model, output is more volatile than
in the standard model and about as volatile as in the
indivisible labor model. The standard deviation of
hours relative to productivity has increased
considerably compared to the standard model, to
We have presented several extensions to the standard
real business cycle model and analyzed the extent to
which they help account for the U.S. business cycle
facts, especially those facts concerning hours and productivity. Introducing nonseparable leisure, indivisible
labor, or home production increases the elasticity of
hours worked with respect to short-run productivity
changes. Introducing a second shock, either to government spending or to the home production function, reduces the correlation between hours worked and productivity.19
Note that our goal has not been to convince you that
any of these models is unequivocally to be preferred.
Our goal has been simply to explain some commonly
used real business cycle models and compare their
implications for the basic labor market facts.
16
The two parameters ␪ and ␩ can be calibrated to match the observed
average levels of market capital (producer durables and nonresidential
structures) and home capital (consumer durables and residential structures)
in the U.S. economy. This requires a lower value for ␪ and a higher value
for ␩ than used here, as discussed in Greenwood, Rogerson, and Wright
1992.
17
The exact results are somewhat sensitive to changes in the parameters
e and ␥, for reasons discussed in the next paragraph.
18
Even if the variance of the shock to the home technology is set to
zero, shocks to the market technology will still induce relative productivity
differentials across sectors. And even if the two shocks are perfectly correlated
and of the same magnitude, agents will still have an incentive to switch
between sectors over time because capital is produced exclusively in the
market. It is these effects that are behind the increase in the labor supply
elasticity.
19
Other extensions not considered here can also affect the implications
of the model for the labor market facts, including distorting taxation as in
Braun 1990 or McGrattan 1991 and nominal contracting as in Cho and
Cooley 1990.
11
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
178
SOME EXTENSIONS
References
Backus, David K., and Kehoe, Patrick J.Forthcoming. International
evidence on the historical properties of business cycles. American
Economic Review.
Benhabib, Jess; Rogerson, Richard; and Wright, Randall. 1991. Homework
in macroeconomics: Household production and aggregate
fluctuations. Journal of Political Economy 99 (December): 1166–87.
Blackburn, Keith, and Ravn, Morten O. 1991. Contemporary
macroeconomic fluctuations: An international perspective. Memo
1991–12. University of Aarhus Center for International
Economics.
Braun, R.Anton. 1990. The dynamic interaction of distortionary taxes
and aggregate variables in postwar U.S. data. Working Paper.
University of Virginia.
Cho, Jang-Ok, and Cooley, Thomas F. 1989. Employment and hours
over the business cycle. Working Paper 132. Rochester Center for
Economic Research. University of Rochester.
_______. 1990. The business cycle with nominal contracts. Working
Paper 260. Rochester Center for Economic Research. University
of Rochester.
Christiano, Lawrence J., and Eichenbaum, Martin. 1992. Current realbusiness-cycle theories and aggregate labor-market fluctuations.
American Economic Review 82 (June): 430–50.
Greenwood, Jeremy; Hercowitz, Zvi; and Huffman, Gregory W. 1988.
Investment, capacity utilization and the real business cycle. American
Economic Review 78 (June): 402–17.
Greenwood, Jeremy; Rogerson, Richard; and Wright, Randall. 1992.
Household production in real business cycle theory. Manuscript.
University of Western Ontario.
Hansen, Gary D. 1985. Indivisible labor and the business cycle. Journal
of Monetary Economics 16 (November): 309–27.
_______. 1989. Technical progress and aggregate fluctuations.
Department of Economics Working Paper 546. University of
California, Los Angeles.
_______. 1991. The cyclical and secular behavior of the labor input:
Comparing efficiency units and hours worked. Manuscript.
University of California, Los Angeles.
Hansen, Gary D., and Prescott, Edward C. 1991. Recursive methods for
computing equilibria of business cycle models. Discussion Paper
36. Institute for Empirical Macroeconomics (Federal Reserve Bank
of Minneapolis).
Hotz, V.Joseph; Kydland, Finn E.; and Sedlacek, Guilherme L. 1988.
Intertemporal preferences and labor supply. Econometrica 56
(March): 335–60.
Juster, F.Thomas, and Stafford, Frank P. 1991. The allocation of time:
Empirical findings, behavioral models, and problems of
measurement. Journal of Economic Literature 29 (June): 471–522.
King, Robert G.; Plosser, Charles I.; and Rebelo, Sergio T. 1987. Production,
growth and cycles: Technical appendix . Manuscript. University
of Rochester.
Kydland, Finn E., and Prescott, Edward C. 1982. Time to build and
aggregate fluctuations. Econometrica 50 (November): 1345–70.
_______. 1990. Business cycles: Real facts and a monetary myth. Federal
Reserve Bank of Minneapolis Quarterly Review 14 (Spring): 3–18.
_______. 1991. Hours and employment variation in business cycle theory.
Economic Theory 1:63–81.
McGrattan, Ellen R. 1991. The macroeconomic effects of distortionary
taxation. Discussion Paper 37. Institute for Empirical
Macroeconomics (Federal Reserve Bank of Minneapolis).
Prescott, Edward C. 1986. Theory ahead of business cycle measurement
Federal Reserve Bank of Minneapolis Quarterly Review 10 (Fall): 9–22.
Rogerson, Richard. 1984. Topics in the theory of labor markets. Ph.D.
dissertation. University of Minnesota.
_______. 1988. Indivisible labor, lotteries and equilibrium. Journal of
Monetary Economics 21 (January): 3–16.
Rogerson, Richard, and Wright, Randall. 1988. Involuntary
unemployment in economies with efficient risk sharing. Journal of
Monetary Economics 22 (November): 501–15.
Shell, Karl, and Wright, Randall. Forthcoming. Indivisibilities, lotteries
and sunspot equilibria. Economic Theory.
Wright, Randall. 1991. The labor market implications of unemployment
insurance and short-time compensation. Federal Reserve Bank of
Minneapolis Quarterly Review 15 (Summer): 11–19.
12
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
CHAPTER 10
179
Current Real-Buisness-Cycle Theories and
Aggregate Labor-Market Fluctuations
By LAWRENCE J.CHRISTIANO AND MARTIN EICHENBAUM*
Hours worked and the return to working are weakly correlated. Traditionally, the ability
to account for this fact has been a litmus test for macroeconomic models. Existing realbusiness-cycle models fail this test dramatically. We modify prototypical real-business-cycle
models by allowing government consumption shocks to influence labor-market dynamics.
This modification can, in principle, bring the models into closer conformity with the data.
Our empirical results indicate that it does. (JEL E32, C12, C52, C13, C51)
In this paper, we assess the quantitative
implications of existing real-business-cycle
(RBC) models for the time-series properties
of average productivity and hours worked.
We find that the single most salient shortcoming of existing RBC models lies in their
predictions for the correlation between
these variables. Existing RBC models predict that this correlation is well in excess of
0.9, whereas the actual correlation is much
closer to zero.1 This shortcoming leads us
to add to the RBC framework aggregate demand shocks that arise from stochastic
movements in government consumption.
According to our empirical results, this
change substantially improves the models’
empirical performance.
The ability to account for the observed
correlation between the return to working and
the number of hours worked has traditionally
been a litmus test for aggregate economic
models. Thomas J.Sargent (1987 p. 468), for
example, states that one of the primary
empirical patterns casting doubt on the
classical and Keynesian models has been the
observation by John T.Dunlop (1938) and
Lorie Tarshis (1939) “alleging the failure of
real wages to move countercyclically.” The
classical and Keynesian models share the
assumption that real wages and hours worked
lie on a stable, downward-sloped marginal
productivity-of-labor curve. 2 Consequently,
they both counterfactually predict a strong
negative correlation between real wages and
hours worked. Modern versions of what
Sargent (1987 p. 468) calls the “Dunlop-Tarshis
observation” continue to play a central role in
assessing the empirical plausibility of different
business-cycle models.3 In discussing Stanley
Fischer’s (1977) sticky-wage business-cycle
*Christiano: Federal Reserve Bank of Minneapolis, Minneapolis,
MN 55480; Eichenbaum: Northwestern University, Evanston, IL 60208,
National Bureau of Economic Research, and Federal Reserve Bank of
Chicago. This paper is a substantially revised version of NBER Working Paper No. 2700, “Is Theory Really Ahead of Measurement? Current
Real Business Cycle Theories and Aggregate Labor Market Fluctuations.” We thank Rao Aiyagari, Paul Gomme, Finn Kydland, Ed Prescott,
and Mark Watson for helpful conversations. Any views expressed here
are ours and not necessarily those of any part of the Federal Reserve
System.
1
This finding is closely related to Bennett McCallum’s (1989) observation that existing RBC models generate grossly counterfactual predictions for the correlation between average productivity and output.
2
As John Maynard Keynes (1935 p. 17) says, “…I am not disputing
this vital fact which the classical economists have (rightly) asserted as
indefeasible. In a given state of organisation, equipment and technique, the
real wage earned by a unit of labour has a unique (inverse) correlation
with the volume of employment.”
3
For example, Robert J.Barro and Herschel I.Grossman (1971) cite
the Dunlop-Tarshis observation to motivate their work on disequilibrium
theories. Also, Edmund S.Phelps and Sidney G.Winter, Jr. (1970 p. 310)
and Franco Modigliani (1977 p. 7) use this observation to motivate their
work on noncompetitive approaches to macroeconomics. Finally, Robert
E.Lucas, Jr. (1981 p. 13) cites the Dunlop-Tarshis observation in motivating his work on capacity and overtime.
430
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
180
SOME EXTENSIONS
VOL. 82 NO. 3
CHRISTIANO AND EICHENBAUM: CURRENT RBC THEORIES
model, for example, Bennett McCallum (1989
p. 191) states that
…the main trouble with the Fischer
model concerns its real wage behavior.
In particular, to the extent that the model
itself explains fluctuations in output and
empløyment, these should be inversely
related to real wage movements: output
should be high, according to the model,
when real wages are low. But in the
actual U.S. economy there is no strong
empirical relation of that type.
In remarks particularly relevant to RBC models, Robert E.Lucas (1981 p. 226) says that
“observed real wages are not constant over the
cycle, but neither do they exhibit consistent
pro- or countercyclical tendencies. This suggests that any attempt to assign systematic real
wage movements a central role in an explanation of business cycles is doomed to failure.”
Existing RBC models fall prey to this (less wellknown) Lucas critique. Unlike the classical and
Keynesian models, which understate the correlation between hours worked and the return
to working, existing RBC models grossly overstate that correlation. According to existing
RBC models, the only impulses generating fluctuations in aggregate employment are stochastic
shifts in the marginal product of labor. Loosely
speaking, the time series on hours worked and
the return to working are modeled as the intersection of a stochastic labor-demand curve
4
Although Finn E.Kydland and Edward C.Prescott (1982) and
Prescott (1986) never explicitly examine the hours/real-wage correlation
implication of existing RBC models, Prescott (1986 p. 21) does implicitly
acknowledge that the failure to account for the Dunlop-Tarshis observation is the key remaining deviation between economic theory and observations: “The key deviation is that the empirical labor elasticity of output is
less than predicted by theory.” To see the connections, denote the empirical
431
with a fixed labor-supply curve. Not surprisingly, therefore, these theories predict a strong
positive correlation between hours worked and
the return to working.4
Several strategies exist for modeling the
observed weak correlation between measures
of these variables. One is to consider models
in which the return to working is unaffected
by shocks to agents’ environments, regardless
of whether the shocks are to aggregate demand
or to aggregate supply. Pursuing this strategy,
Olivier Jean Blanchard and Stanley Fischer
(1989 p. 372) argue that the key assumption of
Keynesian macro models—nominal wage and
price stickiness—is motivated by the view that
aggregate demand shocks affect employment
but not real wages. Another strategy is simply
to abandon one-shock models of aggregate
fluctuations and suppose that the business
cycle is generated by a variety of impulses.
Under these conditions, the Dunlop-Tarshis
observation imposes no restrictions per se on
the response of real wages to any particular
type of shock. Given a specific structural model,
however, it does impose restrictions on the
relative frequency of different types of shocks.
This suggests that one strategy for reconciling
existing RBC models with the Dunlop-Tarshis
observation is to find measurable economic
impulses that shift the labor-supply function.5
With different impulses shifting the laborsupply and labor-demand functions, there is
no a priori reason for hours worked to be
labor elasticity by ␩. By definition, ␩≡corr(y, n)σy/σn, where corr(i, j) is the
correlation between i and j, σi is the standard deviation of i, y is the logarithm of detrended output, and n is the logarithm of hours. Simple arithmetic yields corr(y-n, n)=[η-1](␴n/␴y-n). If, as Prescott claims, RBC models
do well at reproducing the empirical estimates of ␴n/␴y-n, then saying that
the models overstate ␩ is equivalent to saying that they overstate corr(y-n,
n). In Prescott’s model and with his assumed market structure, corr(y-n, n)
is exactly the same as the correlation between real wages and hours worked.
(Also, under log detrending, y-n is log detrended productivity.)
5
An alternative strategy is pursued by Valerie R.Bencivenga (1992),
who allows for shocks to labor suppliers’ preferences. Matthew D.Shapiro
and Mark W.Watson (1988) also allow for unobservable shocks to the
labor-supply function. Jess Benhabib et al. (1991) and Jeremy Greenwood
and Zvi Hercowitz (1991) explore the role of shocks to the home production technology.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
AGGREGATE LABOR MARKET FLUCTUATIONS
432
THE AMERICAN ECONOMIC REVIEW JUNE
correlated in any particular way with the return
to working.
Candidates for such shocks include tax rate
changes, innovations to the money supply,
demographic changes in the labor force, and
shocks to government spending. We focus on
the last of these. By ruling out any role for
government-consumption shocks in labormarket dynamics, existing RBC models
implicitly assume that public and private
consumption have the same impact on the
marginal utility of private spending. Robert
J.Barro (1981, 1987) and David Alan Aschauer
(1985) argue that if $1 of additional public
consumption drives down the marginal utility
of private consumption by less than does $1 of
additional private consumption, then positive
shocks to government consumption in effect
shift the labor-supply curve outward. With
diminishing labor productivity, but without
technology shocks, such impulses will generate
a negative correlation between hours worked
and the return to working in RBC models.
In our empirical work, we measure the
return to working by the average productivity
of labor rather than real wages. We do this for
both empirical and theoretical reasons. From
an empirical point of view, our results are not
very sensitive to whether the return to
working is measured by real wages or average
productivity: Neither displays a strong positive
correlation with hours worked, so it seems
appropriate to refer to the low correlation
between the return to working and hours
worked as the Dunlop-Tarshis observation,
regardless of whether the return to working
is measured by the real wage or average
productivity. From a theoretical point of view,
a variety of ways exist to support the quantity
allocations emerging from RBC models. By
using average productivity as our measure of
the return to working, we avoid imposing the
assumption that the market structure is one
in which real wages are equated to the marginal
product of labor on a period-by-period basis.
Also, existing parameterizations of RBC
models imply that marginal and average
productivity of labor are proportional to each
181
1992
other. For the calculations we perform, the
two are interchangeable.
Our empirical results show that
incorporating government into the analysis
substantially improves the RBC models’
performance. Interestingly, the impact of this
perturbation is about as large as allowing for
nonconvexities in labor supply of the type
stressed by Gary D.Hansen (1985) and
Richard Rogerson (1988). Once government
is incorporated into the analysis, we cannot
reject the hypothesis that a version of the
Hansen-Rogerson indivisible-labor model is
consistent with both the observed correlation
between hours worked and average
productivity and the observed volatility of
hours worked relative to average productivity.
This is not true if government is excluded
from the analysis.
The paper is organized as follows. In
Section I, we describe a general equilibrium
model that nests as special cases a variety of
existing RBC models. In Section II, we present
our econometric methodology for estimating
and evaluating the empirical performance of
the model. In Section III, we present our
empirical results. In Section IV, we offer some
concluding remarks.
I. Two Prototypical Real-Business-Cycle Models
In this section, we present two prototypical
real-business-cycle models. One is a stochastic
version of the one-sector growth model
considered by Kydland and Prescott (1980 p.
174). The other is a version of the model
economy considered by Hansen (1985) in
which labor supply is indivisible. In both of
our models, we relax the assumption implicit
in existing RBC models that public and private
spending have identical effects on the marginal
utility of private consumption.
We make the standard RBC assumption
that the time series on the beginning-of-periodt per capita stock of capital (kt), private time-t
and hours worked at time t
consumption
(nt) correspond to the solution of a socialplanning problem which can be decentralized
as a Pareto-optimal competitive equilibrium.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
182
SOME EXTENSIONS
VOL. 82 NO. 3
CHRISTIANO AND EICHENBAUM: CURRENT RBC THEORIES
The following problem nests both our models
as special cases. Let N be a positive scalar that
denotes the time-t endowment of the
representative consumer, and let γ be a positive
scalar. The social planner ranks streams of
consumption services (ct), leisure (N-nt), and
publicly provided goods and services (gt)
according to the criterion function
(1)
Following Barro (1981, 1987), Roger C.
Kormendi (1983), and Aschauer (1985), we
suppose that consumption services are related
to private and public consumption as follows:
(2)
where α is a parameter that governs the sign
and magnitude of the derivative of the
with respect to gt.6
marginal utility of
Throughout, we assume that agents view gt as
an uncontrollable stochastic process. In
addition, we suppose that gt does not depend
on the current or past values of the
endogenous variables in the model.7
We consider two specifications for the
function V(·). In the divisible-labor model, V(·) is
given by
(3)
433
(4)
This specification can be interpreted in at least
two ways. One is that the specification simply
reflects the assumption that individual utility
functions are linear in leisure. The other interpretation builds on the assumption that labor supply is indivisible. Under this second
interpretation, individuals can either work
some positive number of hours or not work at
all. Assuming that agents’ utility functions are
separable across consumption and leisure,
Rogerson (1988) shows that a market structure in which individuals choose lotteries
rather than hours worked will support a Paretooptimal allocation of consumption and leisure.
The lottery determines whether individuals
work or not. Under this interpretation, (4)
represents a reduced-form preference-ordering that can be used to derive the Pareto-optimal allocation by solving a fictitious social-planning problem. This is the specification used
by Hansen (1985).
Per capita output yt is produced using the
Cobb-Douglas production function
(5)
where 0<␪<1 and zt is an aggregate shock to
technology that has the time-series representation
(6)
In the indivisible-labor model, V(·) is given by
Here ␭t is a serially uncorrelated independent
and identically distributed process with mean
␭ and standard error ␴␭. The aggregate resource constraint is given by
(7)
6
We can generalize the criterion function (1) by writing it as
ln(ct)+␥V(N-nt)+φ (gt), where φ (·) is some positive concave function. As
long as gt is modeled as an exogenous stochastic process, the presence of
such a term has no impact on the competitive equilibrium. However, the
presence of φ (gt)>0 means that agents do not necessarily feel worse off
when gt is increased. The fact that we have set φ (·)≡0 reflects our desire to
minimize notation, not the view that the optimal level of gt is zero.
7
Under this assumption, gt is isomorphic to an exogenous shock to
preferences and endowments. Consequently, existing theorems which establish that the competitive equilibrium and the social-planning problem
coincide are applicable.
That is, per capita consumption and
investment cannot exceed per capita output.
At time 0, the social planner chooses
to
contingency plans for
maximize (1) subject to (3) or (4), (5)–(7), k0,
and a law of motion for gt. Because of the
nonsatiation assumption implicit in (1), we can,
without loss of generality, impose strict equality
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
AGGREGATE LABOR MARKET FLUCTUATIONS
434
THE AMERICAN ECONOMIC REVIEW JUNE
183
1992
in (7). Substituting (2), (5), and this version of
(7) into (1), we obtain the following social-planning
problem: maximize subject to k0, a law of motion
same as those that permanently enhance the
economy’s productive ability.
Substituting (9) into (8), we obtain the
criterion function:
(8)
(11)
where
(12)
for gt, and V(·) given by either (3) or (4), by
choice of contingency plans for {kt+1, nt: t≥0}.
It is convenient to represent the socialplanning problem (8) in a way such that all of
the planner’s decision variables converge in
nonstochastic steady state. To that end, we
define the following detrended variables:
(9)
To complete our specification of agents’ environment, we assume that g-t evolves according
(10)
to where
is the mean of
and µt is the innovation in
with standard deviation ␴µ. Notice that gt has two components, zt and . Movements in zt produce
permanent changes in the level of government
consumption, whereas movements in g- t produce temporary changes in gt. With this specification, the factors giving rise to permanent
shifts in government consumption are the
and where
and V(·) is
given by either (3) or (4). Consequently, the original planning problem is equivalent to the prob(10), and
lem of maximizing (11), subject to
(12), and V(·) is given by either (3) or (4). Since ␬
is beyond the planner’s control, it can be disregarded in solving the planner’s problem.
The only case in which an analytical
solution for this problem is possible occurs
when ␣=␦=1 and the function V( ·) is given
by (3). John B.Long, Jr., and Charles I.Plosser
(1983) provide one analysis of this case.
Analytical solutions are not available for general
values of ␣ and ␦. We use Christiano’s (1988)
log-linear modification of the procedure used
by Kydland and Prescott (1982) to obtain an
approximate solution to our social-planning
problem. In particular, we approximate the
optimal decision rules with the solution to the
linear-quadratic problem obtained when the
function r in (12) is replaced by a function R,
which is quadratic in ln(nt),
and ␭t. The function R is the secondorder Taylor expansion of r[exp(A1), exp(A2),
exp(A3), exp(A4), A5] about the point
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
184
SOME EXTENSIONS
VOL. 82 NO. 3
CHRISTIANO AND EICHENBAUM: CURRENT RBC THEORIES
Here n and denote the steady-state values of
in the nonstochastic version of (11)
nt and
obtained by setting ␴λ=␴µ=0.
Results in Christiano (1988) establish that
the decision rules which solve this problem
are of the form
(13)
and
(14)
In (13) and (14), rk, dk, ek, rn, dn, and en are scalar
functions of the model’s underlying structural
parameters.8
To gain intuition for the role of
in
aggregate labor-market fluctuations, it is useful
to discuss the impact of three key parameters
(␣, ␳, and ␥) on the equilibrium response of nt
to
This response is governed by the
coefficient dn.
First, notice that when ␣=1 the only way
and gt enter into the social planner’s preferences
and constraints is through their sum,
Thus, exogenous shocks to gt induce one-forleaving other
one offsetting shocks in
variables like yt, kt+1, and nt unaffected. This
implies that the coefficients dn and dk in the
planner’s decision rules for kt+1 and nt both equal
zero. Consequently, the absence of a role for gt
in existing RBC models can be interpreted as
reflecting the assumption that ␣=1.
Second, consider what happens when ␣<1.
The limiting case of ␣=0 is particularly useful
for gaining intuition. Government
consumption now is formally equivalent to a
pure resource drain on the economy; agents
respond to an increase in government
consumption as if they had suffered a
reduction in their wealth. (As footnote 6
indicates, this does not imply that they have
8
Christiano (1987a, 1988 footnotes 9, 18) discusses the different
properties of the log-linear approximation that we use here and linear
approximations of the sort used by Kydland and Prescott (1982).
435
suffered a reduction in utility.) The coefficient
dn is positive, since we assume that leisure is a
are
normal good. That is, increases in
associated with increases in nt and decreases
in y t/n t . Continuity suggests that d n is
decreasing in a. The same logic suggests that
dn is an increasing function of ␳, since the wealth
effect of a given shock to is increasing in ␳.
For a formal analysis of the effects of
government consumption in a more general
environment than the one considered here,
see S.Rao Aiyagari et al. (1990).
Finally, consider the impact of ␥ on
aggregate labor-market fluctuations. In
several experiments, we found that en and dn
were increasing in ␥ (for details, see
Christiano and Eichenbaum [1990a]). To
gain intuition into this result, think of a
version of the divisible-labor model in which
the gross investment decision rule is fixed
exogenously. In this simpler model economy,
labor-market equilibrium is the result of the
intersection of static labor-supply and labordemand curves. Given our assumptions
regarding the utility function, the response
of labor supply to a change in the return to
working is an increasing function of ␥ ; that
is, the labor-supply curve becomes flatter as ␥
increases. By itself, this makes the
equilibrium response of nt to ␭t (which shifts
the labor-demand curve) an increasing
function of ␥. This relationship is consistent
with the finding that en is increasing in ␥ in
our model. With respect to d n , it is
straightforward to show that, in the static
framework, the extent of the shift in the
labor-supply curve induced by a change in
is also an increasing function of ␥ . This is
also consistent with the finding that dn is an
increasing function of ␥ in our model.
That en and dn are increasing in ␥ leads us
to expect that the volatility of hours worked
will also be an increasing function of ␥ .
However, we cannot say a priori what impact
larger values of ␥ will have on the DunlopTarshis correlation, because larger values of en
drive that correlation up, but larger values of
dn drive it down.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
AGGREGATE LABOR MARKET FLUCTUATIONS
436
THE AMERICAN ECONOMIC REVIEW JUNE
II. Econometric Methodology
In this section, we describe three things:
our strategy for estimating the structural
parameters of the model and various second
moments of the data, our method for evaluating the model’s implications for aggregate
labor-market fluctuations, and the data used
in our empirical analysis. While similar in
spirit, our empirical methodology is quite
different from the methods typically used
to evaluate RBC models. Much of the existing RBC literature makes little use of formal
econometric methods, either when model
parameter values are selected or when the
fully parameterized model is compared with
the data. Instead, the RBC literature tends
to use a variety of informal techniques, often referred to as calibration. In contrast, we
use a version of Lars Peter Hansen’s (1982)
generalized method-of-moments (G M M)
procedure at both stages of the analysis.
Our estimation criterion is set up so that, in
effect, estimated parameter values equate
model and sample first moments of the
data. It turns out that these values are very
similar to the values used in existing RBC
studies. An important advantage of our
GMM procedures, however, is that they let
us quantify the degree of uncertainty in our
estimates of the model’s parameters. This
turns out to be an important ingredient of
our model-evaluation techniques.
A. Estimation
Now we will describe our estimation strategy.
The parameters of interest can be divided into
two groups. Let ⌿ 1 denote the model’s eight
structural parameters:
(15)
The parameters N, ␤, and a were not estimated.
Instead, we fixed N at 1,369 hours per quarter
and set the parameter ␤ so as to imply a 3percent annual subjective discount rate; that
is, ␤=(1.03)-0.25. Two alternative values of α
were considered: α=0 and α=1.
and
Given estimated values of ⌿ 1,
185
1992
distribution assumptions on µt and ␭t, our
model provides a complete description of the
data-generating process. (Here T denotes the
number of observations in our sample.) This
can be used to compute the second moments
of all the variables of the model. Suppose, for
the time being, that we can abstract from
say, because we
sampling uncertainty in
have a large data sample. Then the second
will coincide with
moments implied by
the second moments of the stochastic process
generating the data only if the model has been
specified correctly.
This observation motivates our strategy
for assessing the empirical plausibility of the
model. First we calculate selected second
moments of the data using our model
. Then we estimate the same
evaluated at
second moments directly, without using the
model. Our test then compares these two sets
of second moments and determines whether
the differences between them can be
accounted for by sampling variation under the
null hypothesis that the model is correctly
specified.
To this end, it is useful to define ⌿ 2 to be
various second moments of the data. Our
dkt, kt, yt, (y/n)t and gt all display
measures of
marked trends, so some stationarity-inducing
transformation of the data must be adopted
for second moments to be well defined. (Here
dk t denotes gross investment.) The
transformation we used corresponds to the
Hodrick and Prescott (H P) detrending
procedure discussed by Robert J.Hodrick and
Prescott (1980) and Prescott (1986). We used
the H P transformation because many
researchers, especially Kydland and Prescott
(1982, 1988), G.Hansen (1985), and Prescott
(1986), have used it to investigate RBC models.
Also, according to our model, the logarithms
dkt, kt, yt, (y/n)t, and gt are all differenceof
stationary stochastic processes. That the HP
filter is a stationarity-inducing transformation
for such processes follows directly from
results of Robert G.King and Sergio T.Rebelo
(1988). We also used the first-difference filter
in our analysis. Since the results are not
substantially different from those reported
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
186
SOME EXTENSIONS
VOL. 82 NO. 3
CHRISTIANO AND EICHENBAUM: CURRENT RBC THEORIES
437
here, we refer the reader to Christiano and
Eichenbaum (1990a) for details. The
parameters in ⌿ 2 are
rate of substitution of goods in consumption
equals the time-t expected value of the marginal
return to physical investment in capital.
Therefore,
(16)
(18)
where σx denotes the standard deviation of the
variable x, with x ={cp, y, dk, n, y/n, g}, and corr(y/
n, n) denotes the correlation between y/n and n.
1. The Unconditional Moments Underlying Our
Estimator of ⌿ 1.—The procedure we used to
estimate the elements of ⌿ 1 can be described
as follows. Our estimator of ␦ is, roughly, the
rate of depreciation of capital implicit in the
empirical capital-stock and investment series.
The estimators of ␪ and ␥ are designed to
allow the model to reproduce the average value
of the capital: output ratio and hours worked
observed in the data. The point estimates of ␳,
and ␴ µ are obtained by applying ordinary
least squares to data on g t/z t, where z t is
constructed using the estimated value of ␪.
Finally, our point estimates of ␭ and ␴␭ are the
mean growth rate of output and the standard
deviation of the growth rate of zt, respectively.
We map these estimators into a G M M
framework to get an estimate of the sampling
distribution of our estimator of ⌿ 1. We need
that estimate for our diagnostic procedures.
To use GMM, we express the estimator of
⌿ 1,
as the solution to the sample analog
of first-moment conditions. We now describe
these conditions. According to our model,
␦ =1+(dk t/k t )-(k t+1 /k t). Let δ* denote the
unconditional mean of the time series [1+(dkt/
kt)-(kt+1/kt)]; that is,
This is the moment restriction that underlies
our estimate of ␪.
The first-order necessary condition for
hours worked requires that, for all t, the
marginal productivity of hours times the
marginal utility of consumption equals the
marginal disutility of working. This implies
the condition ␥ =(1-␪)(yt/nt)/ [ctV’(N-nt)] for all
t. Let ␥ * denote the unconditional expected
value of the time series on the right side of
that condition; that is,
(19)
We identify ␥ with a consistent estimate of
the parameter ␥ *.
Next, consider the random variable
Here ⌬ denotes the first-difference operator.
Under the null hypothesis of balanced growth,
␭=E␭t, the unconditional growth rate of output. Therefore,
(20)
(17)
We identify ␦ with a consistent estimate of the
parameter ␦*.
The social planner’s first-order necessary
condition for capital accumulation requires
that the time-t expected value of the marginal
The relations in (20) summarize the moment
restrictions underlying our estimators of ␭
and ␴ ␭.
Our assumptions regarding the stochastic
process generating government consumption
imply the unconditional moment restrictions:
These moment restrictions can be used to estimate ␳, and ␴µ.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
AGGREGATE LABOR MARKET FLUCTUATIONS
438
THE AMERICAN ECONOMIC REVIEW JUNE
187
1992
struction, H P-filtered data have a zero
mean.
Equations (23) consist of six unconditional
moment restrictions involving the six elements
of ⌿2. These restrictions can be summarized as
(21)
(24)
Equations (17)–(21) consist of eight
unconditional moment restrictions involving
the eight elements of ⌿ 1 . These can be
summarized as
(22)
where
is the true value of ⌿ 1 and H1, t(⌿ 1)
is the 8×1 random vector which has as its
elements the left sides of (17)–(21) before expectations are taken.
2. The Unconditional Moments Underlying Our
Estimator of ⌿ 2.—Our estimator of the elements
of ⌿ 2 coincides with standard second-moment
estimators. We find it convenient to map these
into the GMM framework. The first-moment
conditions we use are
In (23) we have used the fact that, by con(23)
where
is the true value of ⌿ 2 and H2; t(⌿ 2)
is the 6×1 vector-valued function which has
as its elements the left sides of (23) before
expectations are taken.
In order to discuss our estimator, it is
convenient to define the 14×1 parameter vector
⌿ =[⌿
⌿ 1 ⌿ 2 ]′ and the 14×1 vector-valued
function
With this
notation, the unconditional moment
restrictions (22) and (24) can be written as
(25)
where
the vector of true values of ⌿ . Let gT denote the 14×1 vector-valued function
(26)
⌿ 0) is a stationary
Our model implies that Ht(⌿
and ergodic stochastic process. Since gT(·) has
the same dimension as ⌿ , it follows from
L.Hansen (1982) that the estimator ⌿ T, defined by the condition
0, is consistent for ⌿ 0.
Let D T denote the matrix of partial
derivatives
(27)
⌿ T. It then follows from reevaluated at ⌿ =⌿
sults in L.Hansen (1982) that a consistent estimator of the variance-covariance matrix of
is given by
T
(28)
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
188
SOME EXTENSIONS
VOL. 82 NO. 3
CHRISTIANO AND EICHENBAUM: CURRENT RBC THEORIES
Here, ST is a consistent estimator of the spectral density matrix of Ht(⌿0) at frequency
zero.9
B. Testing
Now we describe how a Wald-type test statistic described in Eichenbaum et al. (1984) and
Whitney K.Newey and Kenneth D.West
(1987) can be used to assess formally the plausibility of the model’s implications for subsets
of the second moments of the data. Our empirical analysis concentrates on assessing the
model’s implications for the labor-market moments, [corr(y/n), ␴n/␴y/n].10 Here, we describe
our procedure for testing this set of moments,
a procedure which can be used for any finite
set of moments.
Given a set of values for ⌿1, our model
implies particular values for [corr(y/n), ␴n/␴y/
n ] in population. We represent this
relationship by the function f that maps
into :
(29)
439
Here, f1(⌿1) and f2(⌿2) denote the model’s
implication for corr(y/n, n) and ␴n/␴y/n in population, conditional on the model parameters,
⌿1. The function f(·) is highly nonlinear in
⌿1 and must be computed using numerical
methods. We use the spectral technique described in Christiano and Eichenbaum
(1990b).
Let A be the 2×14 matrix composed of
zeros and ones with the property that
(30)
and let
(31)
Under the null hypothesis that the model is
correctly specified,
(32)
If our data sample were large, then
and (32) could be tested by simply comwith a 2×1 vector of zeros. Howparing
ever,
need not be zero in a small sample,
To test
because of sampling uncertainty in
(32), then, we need the distribution of
under the null hypothesis. Taking a first-order Taylor-series approximation of
about ⌿0 yields
(33)
9
Let
denote the
spectral density matrix of Ht(⌿0) at frequency zero. Our estimator of S0,
ST, uses the damped, truncated covariance estimator discussed by
Eichenbaum and Hansen (1990). The results we report were calculated
by truncating after six lags. Strictly speaking, HP-filtered data do not
satisfy the Eichenbaum and Hansen (1990) assumption that S0 be
nonsingular. This is because our model implies that data need to be
differenced only once to induce stationarity, while the results of King and
Rebelo (1988) show that the HP filter differences more than once. We
think this is not a serious problem from the perspective of asymptoticdistribution theory. This is because our numerical results would have
been essentially unchanged had we worked with a version of the HP
filter in which the extra unit roots were replaced by roots arbitrarily close
to 1. Then, the Eichenbaum and Hansen (1990) analysis would apply
without caveat. What the small-sample properties are in the presence of
unit roots in the data-generating process remains an open and interesting
question.
10
Our formal test does not include ␴n/␴y because this is an exact function
of [corr(y/n, n), ␴n/␴y/n]. To see this, let b=␴n/␴y/n and c=corr(y/n, n). Then,
after some algebraic manipulation,
␴n / ␴y = b / ( 1 + 2 c b + b 2 ) 1/2 .
It follows that a consistent estimator of the
is given
variance-covariance matrix of
by
(34)
An implication of results in Eichenbaum et al.
(1984) and Newey and West (1987) is that the
test statistic
(35)
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
AGGREGATE LABOR MARKET FLUCTUATIONS
440
THE AMERICAN ECONOMIC REVIEW
is asymptotically distributed as a chi-square
random variable with two degrees of freedom.
We used this fact to test the null hypothesis
(32).
C. The Data
Next we describe the data used in our analysis. In all of our empirical work, private conwas measured as quarterly real
sumption,
expenditures on nondurable consumption
goods plus services plus the imputed service
flow from the stock of durable goods. The
first two time series came from the U.S. Department of Commerce’s Survey of Current Business (various issues). The third came from the
data base documented in Flint Brayton and
Eileen Mauskopf (1985). Government consumption, gt, was measured by real government purchases of goods and services minus
real investment of government (federal, state,
and local).11
A measure of government investment was
provided to us by John C.Musgrave of the
U.S. Bureau of Economic Analysis. This
measure is a revised and updated version of
the measure discussed in Musgrave (1980).
Gross investment, dkt, was measured as privatesector fixed investment plus government fixed
investment plus real expenditures on durable
goods.
The capital-stock series, kt, was chosen to
match the investment series. Accordingly, we
measured k t as the sum of the stock of
consumer durables, producer structures and
equipment, government and private residential
capital, and government nonresidential capital.
plus
Gross output, yt, was measured as
gt plus dkt plus time-t inventory investment.
Given our consumption series, the difference
between our measure of gross output and the
one reported in the Survey of Current Business is
189
JUNE 1992
that ours includes the imputed service flow
from the stock of consumer durables but
excludes net exports.
We used two different measures of hours
worked and average productivity. Our first
measure of hours worked corresponds to the
one constructed by G.Hansen (1984) which is
based on the household survey conducted by
the U.S. Department of Labor. A
corresponding measure of average productivity
was constructed by dividing our measure of
gross output by this measure of hours. For
convenience, we refer to this measure of nt
and (y/n)t as household hours worked and
household productivity.
A potential problem with our measure of
household average productivity is that gross
output covers more sectors than does the
household hours data (for details, see appendix
1 of Christiano and Eichenbaum [1988]). In
order to investigate the quantitative impact of
this problem, we considered a second measure
of hours worked and productivity which
covers the same sectors: output per hour of all
persons in the nonagricultural business sector
(CITIBASE mnemonic LBOUTU) and per
capita hours worked by wage and salary
workers in private nonagricultural
establishments as reported by the U.S.
Department of Labor (Bureau of Labor
Statistics, IDC mnemonic HRSPST). For
convenience, we refer to this measure of nt
and (y/n)t as establishment hours worked and
establishment productivity.
All data, except those for (y/n) t, were
converted to per capita terms using an
efficiency-weighted measure of the population.
The data cover the period from the third
quarter of 1955 through the fourth quarter of
1983 (1955:3–1983:4) (for further details on
the data, see Christiano [1987b, 1988]).
III. Empirical Results
11
It would be desirable to include in gt a measure of the service flow
from the stock of government-owned capital, since government capital is
included in our measure of kt. Unfortunately, we know of no existing
measures of that service flow. This contrasts with household capital, for
which there are estimates of the service flow from housing and the stock of
consumer durables. The first is included in the official measure of consumption of services, and the second is reported by Brayton and Mauskopf
(1985).
In this section, we report our empirical results. Subsection A discusses the results obtained using the household data while Subsection B presents results based on the establishment data. In each case, our results
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
190
SOME EXTENSIONS
VOL. 82 NO. 3
CHRISTIANO AND EICHENBAUM: CURRENT RBC THEORIES
441
TABLE 1—MODEL PARAMETER E STIMATES (AN D STAN DARD E RRORS) G ENERATED BY
THE HOUSEHOLD DATA SET
Notes: Standard errors are reported in parentheses only for estimated parameters. Other parameters were set a
priori.
are presented for four models. These correspond to versions of the model in Section II
with V given by (3) or (4) and ␣=1 or 0. We
refer to the model with V given by (3) and
␣=1 as our base model.
A. Results for the Household Data
Table 1 reports our estimates of ⌿1 along with
standard errors for the different models. (We
report the corresponding equilibrium decision
rules in Christiano and Eichenbaum [1990a].)
Table 2 documents the implications of our estimates of ⌿1 for various first moments of the
data. To calculate these, we used the fully parameterized models to simulate 1,000 time series, each of length 113 (the number of observations in our data set). First moments were
calculated on each synthetic data set. Table 2
reports the average value of these moments
across synthetic data sets as well as estimates
of the corresponding first moments of the data.
As can be seen, all of the models do
extremely well on this dimension. This is
not surprising, given the nature of our
estimator of ⌿ 1 . Notice that the models
k t,
predict the same mean growth rate for
g t, and y t . This prediction reflects the
balanced-growth properties of our models.
This prediction does not seem implausible
given the point estimates and standard errors
reported in Table 2.12 The models also pre12
The large standard error associated with our estimate of the growth
rate of gt may well reflect a break in the data around 1970. For example,
the sample
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
AGGREGATE LABOR MARKET FLUCTUATIONS
442
THE AMERICAN ECONOMIC REVIEW
191
JUNE 1992
TABLE 2—SELECTED FIRST-MOMENT PROPERTIES, HOUSEHOLD DATA SET
Notes: Numbers in the columns under the “model” heading are averages, across 1,000 simulated data sets, each with 113 observations, of the sample
average of the variables in the first column. Numbers in parentheses are the standard deviations across data sets. The last column reports empirical
averages, with standard errors in parentheses.
dict that the unconditional growth rate of nt
will be zero. This restriction also seems
reasonably consistent with the data.
Table 3 displays estimates of a subset of the
second moments of the household data, as
well as the analog model predictions. All of the
models do reasonably well at matching the
␴dk/␴y, ␴g/␴y, and
estimated values of
␴y. Interestingly, introducing government into
the analysis (i.e., moving from ␣=1 to ␣=0)
actually improves the performance of the
average of the growth rate of gt between 1955:2 and 1970:1 is 0.0060,
whereas between 1970:1 and 1984:1 it is –0.0018.
models with respect to
␴dk/␴y, and ␴g/
␴y but has little impact on their predictions for
␴y. The models do not do well, however, at
matching the volatility of hours worked
relative to output (␴n/␴y). Not surprisingly,
incorporating government into the analysis
(␣=0) generates additional volatility in nt, as
does allowing for indivisibilities in labor supply.
Indeed, the quantitative impact of these two
perturbations to the base model (divisible labor
with ␣=1) is similar. Nevertheless, even when
both effects are operative, the model still
underpredicts the volatility of nt relative to yt.
Similarly, allowing for non-convexities in labor
supply and introducing government into the
analysis improves the model’s performance
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
192
SOME EXTENSIONS
VOL. 82 NO. 3
CHRISTIANO AND EICHENBAUM: CURRENT RBC THEORIES
443
TABLE 3—SECOND-MOMENT PROPERTIES AFTER HP-DETRENDING MODELS ESTIMATED USING THE
HOUSEHOLD DATA SET
Notes: All of the statistics in this table are computed after first logging and then detrending the data using the Hodrick-Prescott filter. Here, ␴i is the standard
deviation of variable i detrended in this way and corr(x, w) is the correlation between detrended x and detrended w. Numbers in the columns under the
“model” heading are averages, across 1,000 simulated data sets, each with 113 observations, of the sample average of the variables in the first column.
Numbers in parentheses are the standard deviations across data sets. The last column reports results for U.S. data with associated standard errors,
computed as discussed in the text, in parentheses.
with respect to the volatility of nt relative to yt/
nt. In fact, the model that incorporates both of
these effects actually overstates the volatility
of nt relative to yt/nt.13
Next we consider the ability of the
different models to account for the DunlopTarshis observation. Table 3 shows that the
prediction of the base model is grossly
13
These results differ in an important way from those of G.Hansen
(1985). Using data processed with the HP filter, he reports that the indivisible labor model with ␣=1 implies a value of ␴n/␴y/n equal to 2.7 (Hansen,
1985 table 1). This is more than twice the corresponding empirical quantity. Our version of this model (␣=1) underpredicts ␴n/␴y/n by more than 20
percent. The reason for the discrepancy is that Hansen models innovations
to technology as having a transient effect on zt, whereas we assume that
their effect is permanent. Consequently, the intertemporal substitution effect
of a shock to technology is considerably magnified in Hansen’s version of
the model.
inconsistent with the observed correlation
between average productivity and hours
worked. Introducing nonconvexities in labor
supply has almost no impact on the model’s
prediction for this correlation.14 However,
introducing government into the analysis
(␣ =0) does reduce the prediction some, at
least moving it in the right direction. But
not nearly enough: the models with ␣=0
still substantially overstate the correlation
14
To gain intuition into this result, consider a static version of our
model, with no capital, in which the wage is equated to the marginal
product of labor in each period. In that model, introducing indivisibilities
can be thought of as flattening the labor-supply schedule, thereby increasing the fluctuations of hours worked relative to the wage. However, as long
as the only shocks are to technology, the correlation between hours worked
and the wage will still be strongly negative, regardless of the slope of labor
supply.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
AGGREGATE LABOR MARKET FLUCTUATIONS
444
THE AMERICAN ECONOMIC REVIEW
193
JUNE 1992
TABLE 4—DIAGNOSTIC RESULTS WITH THE Two DATA SETS
Notes: All results are based on data detrended by the Hodrick-Prescott filter. The numbers in the “U.S. data” column are point estimates based on U.S.
data for the statistic. The portion of this column in panel A is taken directly from Table 3. The numbers in parentheses are the associated standard-error
estimates. The numbers in the columns under the “model” heading are the probability limits of the statistics implied by the indicated model at its estimated
parameter values; the numbers in parentheses are the standard errors of the discrepancy between the statistic and its associated sample value, reported
in the U.S. data column. This standard error is computed by taking the square root of the appropriate diagonal element of equation (34). The numbers
in brackets are the associated t statistics. The J statistic is computed using equation (35), and the number in braces is the probability that a chi-square
with two degrees of freedom exceeds the reported value of the associated J statistic.
between average productivity and hours
worked.
Panel A in Table 4 reports the results of
implementing the diagnostic procedures
discussed in Section II. The last row of the
panel (labeled “J”) reports the statistic for
testing the joint null hypothesis that the
model predictions for both corr(y/n, n) and
␴n/ ␴y/n are true. As can be seen, this null
hypothesis is overwhelmingly rejected for
every version of the model. Notice also that
the t statistics (given in brackets in the table)
associated with corr(y/n, n) are all larger than
the corresponding t statistics associated with
␴n/␴y/n. This is consistent with our claim that
the single most striking failure of existing
RBC models lies in their implications for the
Dunlop-Tarshis observation, rather than the
relative volatility of hours worked and average
productivity.
B. Results Based on Establishment Data
There are at least two reasons to believe that
the negative correlation between hours
worked and average productivity reported
above is spurious and reflects measurement
error. One potential source of distortion lies
in the fact that gross output covers more
sectors than household hours. The other
potential source of distortion is that household hours data may suffer from classical
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
194
SOME EXTENSIONS
VOL. 82 NO. 3
CHRISTIANO AND EICHENBAUM: CURRENT RBC THEORIES
445
TABLE 5—MODEL PARAMETER ESTIMATES (AND STANDARD ERRORS) G ENERATED BY
THE ESTABLISHMENT DATA SET
Notes: Standard errors are reported (in parentheses) only for estimated parameters. Other parameters were set a
priori.
measurement error. Classical measurement
error in n will bias standard estimates of corr(y/
n, n) downward.
In order to investigate the quantitative
impact of the coverage problem, we redid our
analysis using establishment hours worked and
establishment average productivity. An
important virtue of these measures is that they
cover the same sectors. With these data, the
estimated value of corr(y/n, n) becomes positive:
0.16 with a standard error of 0.08. This result
is consistent with the view that the negative
correlation reported in panel A of Table 4
reflects, in part, coverage problems with the
household data. Interestingly, our estimate of
␴n/␴y/n is also significantly affected by the move
to the new data set. This increases to 1.64 with
a standard error of 0.16. Thus, while the
models’ performance with respect to the
Dunlop-Tarshis observation ought to be
enhanced by moving to the new data set, it
ought to deteriorate with respect to the relative
volatility of hours worked and output per hour.
Therefore, the net effect of the new data set on
overall inference cannot be determined a priori.
To assess the net impact on the models’
performance, we reestimated the structural
parameters and redid the diagnostic tests
discussed in Section II. The new parameter
estimates are reported in Table 5. The data
used to generate these results are the same as
those underlying Table 1, with two exceptions.
One has to do with the calculations associated
with the intratemporal Euler equation, that is,
the third element of Ht(·). Here we used our
new measure of average productivity, which
is actually an index. This measure of average
productivity was scaled so that the sample
mean of the transformed index coincides with
the sample mean of our measure of yt divided
by establishment hours. The other difference
is that, apart from the calculations involving
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
AGGREGATE LABOR MARKET FLUCTUATIONS
446
THE AMERICAN ECONOMIC REVIEW
y t/n t, we measured n t using establishment
hours.
The new second-moment implications,
with the exception of those pertaining to ␴y,
corr(y/n, n), and ␴n/␴y/n, are very similar to those
reported in Table 3. The new values of ␴y are
0.013 (0.0017) and 0.014 (0.002) for the
versions of the divisible-labor model without
government ( ␣ =1) and with government
(␣=0), respectively, and 0.015 (0.0019) and
0.016 (0.002) for the versions of the indivisiblelabor model without and with government.
(Numbers in parentheses are standard
deviations, across synthetic data sets.) The
fact that these values are all lower than those
in Table 3 primarily reflects our finding that
the variance of the innovation to the Solow
residual is lower with the establishment hours
data.
The results of our diagnostic tests are
summarized in panel B of Table 4. Notice that,
for every version of the model, the J statistic
in panel B is lower than the corresponding
entry in panel A. Nevertheless, as long as
government is not included (i.e., when a=1),
both versions of the model are still rejected at
essentially the zero-percent significance level.
However, this is no longer true when
government is added (when ␣= 0). Then, we
cannot reject the indivisible labor model at even
the 15-percent significance level.
To understand these results, we first
consider the impact of the new data set on
inference regarding the correlation between
hours worked and average productivity.
Comparing the ␣=0 models in panels A and B
of Table 4, we see a dramatic drop in the t
statistics (the bracketed numbers there). There
are two principal reasons for this
improvement. The most obvious reason is that
(y/n, n) is positive in the new data set (0.16),
while it is negative in the old data set (-0.20).
In this sense, the data have moved toward the
model. The other reason for the improved
performance is that the new values of
generate a smaller value for corr(y/n, n). For
example, in the indivisible-labor model with
195
JUNE 1992
␣=0, corr(y/n, n) drops from 0.737 to 0.575. In
part, this reflects the new values of and
Consider first. With the household data
set, is 0.96 (after Founding) for all of the
models; with the establishment data set, is
0.98 (after rounding). As we emphasized in
Section I, increases in ρ are associated with
decreases in the correlation between yt/nt and
nt.15 Next consider With the establishment
data, the estimates of γ are consistently larger
than we obtained with the household data.16
For example, in the indivisible-labor model
with
was 0.00374; now
As we noted in Section I, the impact of a change
in γ on corr(y/n, n) cannot be determined a
priori. As it turns out, the increase in
contributes to a drop in these statistics.17
We now examine the impact of the
establishment data on inference regarding the
relative volatility of hours worked and average
productivity. Comparing panels A and B of
Table 4, we see that in all cases but one, the t
statistics rise. In the exceptional case, that is,
the indivisible-labor model with ␣=0, the
change is very small. Three factors influence
the change in these t statistics. First, the point
estimate of ␴ n / ␴ y/n is larger with the
establishment data. Other things equal, this
15
Consistent with this relationship, corr(y/n, n)=0.644 in the indivisible-labor model with ␣=0, when it is evaluated at the parameter values in Table 1 except with ρ set to 0.98.
16
To see why the establishment data set generates a higher value
of it is convenient to concentrate on the divisible-labor model. The
parameter is invariant to which data set or model is used. In practice,
our estimator of is approximately
where c/y denotes the sample average of
and N/n denotes the sample average of N/nt. Obviously, is a decreasing function
of n. The value of n with the household data set is 320.4, and the implied value of n/N is 0.23. With the establishment data set, n=257.7, and
the implied value of n/N is 0.19. Our estimates of γ are different from
the one used by Kydland and Prescott (1982). This is because Kydland
and Prescott deduce a value of γ based on the assumption that n/N=0.33.
In defending this assumption, Prescott (1986 p. 15) says that “[Gilbert
R.] Ghez and [Gary S.] Becker (1975) find that the household allocates
approximately one-third of its productive time to market activities and
two-thirds to nonmarket activities.” We cannot find any statement of this
sort in Ghez and Becker (1975).
17
For example, in the indivisible labor model with ␣=0 evaluated
at the parameter estimates in Table 1, but with γ increased to 0.0046,
corr(y/n, n)=0.684.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
196
SOME EXTENSIONS
VOL. 82 NO. 3
CHRISTIANO AND EICHENBAUM: CURRENT RBC THEORIES
hurts the empirical performance of all the
models, except the indivisible-labor model with
␣=0. Second, these statistics are estimated less
precisely with the establishment data, and this
contributes to a reduction in the t statistics.
Third, the new parameter estimates lead to an
increase in each model’s implied value of ␴n/␴y/n.
For example, the value of ␴n/␴y/n implied by the
indivisible-labor model with ␣=0 rises to 1.437
from 1.348. In part, this reflects the new values
of and When we evaluate the indivisiblelabor model with ␣ =0 at the parameter
estimates in Table 1, with ρ increased to its
Table 5 value of 0.98, the value of ␴n/␴y/n equals
1.396. The analog experiment with γ increases
the value of this statistic to 1.436.
Comparing panels A and B of Table 4, we
see that inference about the importance of the
role of government consumption appears to
hinge sensitively on which data set is used.
On the one hand, the household data suggest
that the role of government consumption is
minimal. This is because both the divisiblelabor and indivisible-labor models are rejected,
regardless of whether ␣=0 or 1. On the other
hand, the establishment data suggest an
important role for government consumption.
While the divisible-labor model is rejected in
both its variants, the indivisible-labor model
cannot be rejected at conventional significance
levels as long as ␣=0.
In Christiano and Eichenbaum (1990a), we
argue that the sensitivity of inference to which
data set is used is resolved once we allow for
classical measurement error in hours worked.
The basic idea is to assume, as in Prescott
(1986), that the measurement errors in the
logarithm of household and establishment
hours worked are uncorrelated over time and
with each other, as well as with the logarithm
of true hours worked. In Christiano and
Eichenbaum (1990a), we show how to estimate
the parameters of the models considered here,
allowing for this kind of measurement error.
In addition, we did the diagnostic tests that
we have discussed in this paper. The main
findings can be briefly summarized as follows.
First, allowing for measurement error, the
447
indivisible-labor model cannot be rejected at
conventional significance levels as long as
government is incorporated into the analysis.
This is true regardless of whether household
or establishment hours data are used. Second,
the divisible-labor model continues to be
rejected for both data sets, regardless of
whether government is included in the
analysis. Therefore, with this model of
measurement error, inference is not sensitive
to which measure of hours worked is used.
Regardless of whether household or
establishment hours data are used,
incorporating government into the analysis
substantially improves the empirical
performance of the indivisible-labor model.
In Christiano and Eichenbaum (1990a), we
also present evidence that the plausibility of the
divisible-labor model with government is
affected by the choice of stationarity-inducing
transformation. In particular, there is
substantially less evidence against that model
with ␣=0 when the diagnostic tests are applied
to growth rates of the establishment hours data
set and measurement error is allowed for.
IV. Concluding Remarks
Existing RBC theories assume that the only
source of impulses to postwar U.S. business
cycles are exogenous shocks to technology. We
have argued that this feature of RBC models
generates a strong positive correlation between
hours worked and average productivity. Unfortunately, this implication is grossly
counterfactual, at least for the postwar United
States. This led us to conclude that there must
be other quantitatively important shocks driving fluctuations in aggregate U.S. output. We
have focused on assessing the importance of
shocks to government consumption. Our results indicate that, when aggregate demand
shocks arising from stochastic movements in
government consumption are incorporated
into the analysis, the model’s empirical performance is substantially improved.
Two important caveats about our empirical
results should be emphasized. One has to do
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
AGGREGATE LABOR MARKET FLUCTUATIONS
448
THE AMERICAN ECONOMIC REVIEW
with our implicit assumption that public and
private capital are perfect substitutes in the
aggregate production function. Some
researchers, including most prominently
Aschauer (1989), have argued that this
assumption is empirically implausible. To the
extent that these researchers are correct, and
to the extent that public-investment shocks
are important, our assumption makes it easier
for our model to account for the DunlopTarshis observation. This is because these
kinds of shocks have an impact on the model
similar to technology shocks, and they
contribute to a positive correlation between
hours worked and productivity. The other
caveat has to do with another implicit
assumption: that all taxes are lump-sum. We
chose this strategy in order to isolate the role
of shocks to government consumption per se.
We leave to future research the important
task of incorporating distortionary taxation
into our framework. How this would affect
our model’s ability to account for the DunlopTarshis observation is not clear. Recent work
by R.Anton Braun (1989) and Ellen
R.McGrattan (1991) indicates that
randomness in marginal tax rates enhances
the model on this dimension. However, some
simple dynamic optimal-taxation arguments
suggest the opposite. Suppose, for example,
that it is optimal for the government to increase
distortionary taxes on labor immediately in
response to a persistent increase in government
consumption. This would obviously reduce
the positive employment effect of an increase
in government consumption. Still, using a
version of our divisible-labor model, V.V.Chari
et al. (1991) show that this last effect is very
small. In their environment, introducing
government into the analysis enhances the
model’s overall ability to account for the
Dunlop-Tarshis observation. In any event, if it
were optimal for the government to increase
taxes with a lag, we suspect that this type of
distortionary taxation would actually enhance
the model’s empirical performance.
REFERENCES
Aiyagari, S.Rao, Christiano, Lawrence J. and
Eichenbaum, Martin, “The Output, Em-
197
JUNE 1992
ployment, and Interest Rate Effects of Government Consumption,” Discussion Paper
No. 25, Institute for Empirical Macroeconomics (Federal Reserve Bank of Minneapolis and University of Minnesota), 1990.
Aschauer, David Alan, “Fiscal Policy and Aggregate Demand,” American Economic Review,
March 1985, 75, 117–27.
_____, “Does Public Capital Crowd Out Private
Capital?” Journal of Monetary Economics, September 1989, 24, 171–88.
Barro, Robert J., “Output Effects of Government
Purchases,” Journal of Political Economy, December 1981, 89, 1086–1121.
_____, “Government Purchases and Public Services,” in Robert J.Barro, Macroeconomics, 2nd
Ed., New York: Wiley, 1987, pp. 307–39.
______, and Grossman, Herschel I., “A General
Disequilibrium Model of Income and Employment,” American Economic Review, March
1971, 61, 82–93.
Bencivenga, Valerie R., “An Econometric Study
of Hours and Output Variation with Preference Shocks,” International Economic Review,
1992 (forthcoming).
Benhabib, Jess, Rogerson, Richard and Wright,
Randall, “Homework in Macroeconomics:
Household Production and Aggregate Fluctuations,” Journal of Political Economy, December 1991, 6, 1166–81.
Blanchard, Olivier Jean and Fischer, Stanley,
Lectures on Macroeconomics, Cambridge, MA:
MIT Press, 1989.
Braun, R.Anton, “The Dynamic Interaction of
Distortionary Taxes and Aggregate Variables
in Postwar U.S. Data,” unpublished manuscript, University of Virginia, 1989.
Brayton, Flint and Mauskopf, Eileen, “The MPS
Model of the United States Economy,” unpublished manuscript, Board of Governors
of the Federal Reserve System, Division of
Research and Statistics, Washington, DC,
1985.
Chari, V.V., Christiano, Lawrence J. and Kehoe,
Patrick J., “Optimal Fiscal Policy in a Business Cycle Model,” Research Department
Working Paper No. 465, Federal Reserve
Bank of Minneapolis, 1991.
Christiano, Lawrence J., (1987a) “Dynamic Properties of Two Approximate Solutions to a
Particular Growth Model,” Research Department Working Paper No. 338, Federal Reserve Bank of Minneapolis, 1987.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
198
SOME EXTENSIONS
VOL. 82 NO. 3
CHRISTIANO AND EICHENBAUM: CURRENT RBC THEORIES
____, (1987b) Technical Appendix to “Why Does
Inventory Investment Fluctuate So Much?”
Research Department Working Paper No.
380, Federal Reserve Bank of Minneapolis,
1987.
____, “Why Does Inventory Investment Fluctuate So Much?” Journal of Monetary Economics,
March/May 1988, 21, 247–80.
____ and Eichenbaum, Martin, “Is Theory Really Ahead of Measurement? Current Real
Business Cycle Theories and Aggregate Labor Market Fluctuations,” National Bureau
of Economic Research (Cambridge, MA)
Working Paper No. 2700, 1988.
____ and _____, (1990a) “Current Real Business
Cycle Theories and Aggregate Labor Market Fluctuations,” Discussion Paper No. 24,
Institute for Empirical Macroeconomics (Federal Reserve Bank of Minneapolis and University of Minnesota), 1990.
___ and ____, (1990b) “Unit Roots in Real GNP:
Do We Know, and Do We Care?” CarnegieRochester Conference Series on Public Policy, Spring
1990, 32, 7–61.
Dunlop, John T., “The Movement of Real and
Money Wage Rates,” Economic Journal, September 1938, 48, 413–34.
Eichenbaum, Martin and Hansen, Lars Peter,
“Estimating Models With Intertemporal Substitution Using Aggregate Time Series Data,”
Journal of Business and Economic Statistics, January 1990, 8, 53–69.
____, ____ and Singleton, Kenneth J., Appendix
to “A Time Series Analysis of Representative Agent Models of Consumption and Leisure Under Uncertainty,” unpublished manuscript, Northwestern University, 1984.
Fischer, Stanley, “Long-Term Contracts, Rational
Expectations, and the Optimal Money Supply Rule,” Journal of Political Economy, February 1977, 85, 191–205.
Ghez, Gilbert R. and Becker, Gary S., The Allocation of Time and Goods Over the Life Cycle, New
York: National Bureau of Economic Research,
1975.
Greenwood, Jeremy and Hercowitz, Zvi, “The
Allocation of Capital and Time Over the
Business Cycle,” Journal of Political Economy,
December 1991, 6, 1188–1214.
Hansen, Gary D., “Fluctuations in Total Hours
Worked: A Study Using Efficiency Units,”
working paper, University of Minnesota,
1984.
449
____, “Indivisible Labor and the Business Cycle,”
Journal of Monetary Economics, November 1985,
16, 309–27.
Hansen, Lars Peter, “Large Sample Properties of
Generalized Method of Moments Estimators,”
Econometrica, July 1982, 50, 1029–54.
Hodrick, Robert J. and Prescott, Edward C.,
“Post-War U.S. Business Cycles: An Empirical Investigation,” unpublished manuscript,
Carnegie Mellon University, 1980.
Keynes, John Maynard, The General Theory of
Employment, Interest and Money, New York:
Harcourt Brace, 1935.
King, Robert G. and Rebelo, Sergio T., “Low
Frequency Filtering and Real Business
Cycles,” unpublished manuscript, University of Rochester, 1988.
Kormendi, Roger C., “Government Debt, Government Spending, and Private Sector Behavior,” American Economic Review, December
1983, 73, 994–1010.
Kydland, Finn E. and Prescott, Edward C., “A
Competitive Theory of Fluctuations and the
Feasibility and Desirability of Stabilization
Policy,” in Stanley Fischer, ed., Rational Expectations and Economic Policy, Chicago: University of Chicago Press, 1980, pp. 169–87.
____ and ____, “Time to Build and Aggregate
Fluctuations,” Econometrica, November 1982,
50, 1345–70.
____ and ____, “The Workweek of Capital and
Its Cyclical Implications,” Journal of Monetary
Economics, March/ May 1988, 21, 343–60.
Long, John B., Jr., and Plosser, Charles I., “Real
Business Cycles,” Journal of Political Economy,
February 1983, 91, 39–69.
Lucas, Robert E., Jr., Studies in Business-Cycle Theory,
Cambridge, MA: MIT Press, 1981.
McCallum, Bennett, Monetary Economics: Theory and
Policy, New York: Macmillan, 1989.
McGrattan, Ellen R., “The Macroeconomic Effects of Distortionary Taxation,” Discussion
Paper No. 37, Institute for Empirical Macroeconomics (Federal Reserve Bank of Minneapolis and University of Minnesota), 1991.
Modigliani, Franco, “The Monetarist Controversy or, Should We Forsake Stabilization
Policies?” American Economic Review, March
1977, 67, 1–19.
Musgrave, John C., “Government-Owned Fixed
Capital in the United States, 1925–79,” Survey of Current Business, March 1980, 60,
33–43.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
AGGREGATE LABOR MARKET FLUCTUATIONS
450
THE AMERICAN ECONOMIC REVIEW
Newey, Whitney K. and West, Kenneth D., “A
Simple, Positive Semi-definite, Heteroskedasticity and Autocorrelation Consistent
Covariance Matrix,” Econometrica, May 1987,
55, 703–8.
Phelps, Edmund S. and Winter, Sidney G., Jr.,
“Optimal Price Policy under Atomistic Competition,” in Edmund S.Phelps, ed.,
Microeconomic Foundations of Employment and Inflation Theory, New York: Norton, 1970, pp.
309–37.
Prescott, Edward C., “Theory Ahead of Business
Cycle Measurement,” Federal Reserve Bank of
Minneapolis Quarterly Review, Fall 1986, 10,
9–22.
199
JUNE 1992
Rogerson, Richard, “Indivisible Labor, Lotteries
and Equilibrium,” Journal of Monetary Economics, January 1988, 21, 3–16.
Sargent, Thomas J., Macroeconomic Theory, 2nd Ed.,
New York: Academic Press, 1987.
Shapiro, Matthew D. and Watson, Mark W.,
“Sources of Business Cycle Fluctuations,”
National Bureau of Economic Research
(Cambridge, MA) Working Paper No. 2589,
1988.
Tarshis, Lorie, “Changes in Real and Money
Wages,” Economic Journal, March 1939, 49,
150–4.
Survey of Current Business, Washington, DC: U.S.
Department of Commerce, various issues.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
200
CHAPTER 11
The Inflation Tax in a Real Business Cycle Model
By THOMAS F.COOLEY AND GARY D.HANSEN*
Money is incorporated into a real business cycle model using a cash-in-advance constraint. The model
economy is used to analyze whether the business cycle is different in high inflation and low inflation
economies and to analyze the impact of variability in the growth rate of money. In addition, the welfare
cost of the inflation tax is measured and the steady-state properties of high and low inflation economies
are compared.
Current controversies in business cycle theory
have much in common with the macroeconomic
debates of the 1960s. Twenty years ago Milton
Friedman and Walter Heller debated the issue
of whether “money matters.” In the ensuing years
the methods of business cycle research have
changed dramatically but the questions have
remained much the same. In particular, the issue of how much money matters is as timely
now as it was when Friedman and Heller discussed it. In this paper we take the question of
whether money matters to mean three things:
does money and the form of the money supply
rule affect the nature and amplitude of the business cycle? how does anticipated inflation affect
the long-run values of macroeconomic variables? and, what are the welfare costs associated
* W.E.Simon Graduate School of Management and Department of Economics, University of Rochester, Rochester, NY 14627 and
Department of Economics, University of California, Los Angeles, CA 90024, respectively.
We would like to acknowledge helpful comments from Steve LeRoy, David I.Levine, Bob
Lucas, Bennett McCallum, Ellen McGrattan,
Seonghwan Oh, Ed Prescott, Kevin Salyer, Tom
Sargent, Bruce Smith, three anonymous referees, and participants in the Northwestern University Summer Research Conference, August
1988. Earlier versions of this paper were titled
“The Inflation Tax and the Business Cycle.”
The first author acknowledges the support of
the John M.Olin Foundation.
with alternative money supply rules? These are
quite different questions and each implies a distinct sense in which money can affect the
economy. Herein we describe a model economy
that can be used to address these sorts of questions. The setting is similar to one suggested by
Robert Lucas (1987) where money is held due
to a cash-in-advance constraint. We use it to provide estimates of the welfare cost of the inflation
tax and to study the effect of anticipated inflation
on the characteristics of aggregate time-series.
Early equilibrium business cycle models
were influenced greatly by the monetarist
tradition and the empirical findings of Milton
Friedman and Anna Schwartz. They were
models where unanticipated changes in the
money supply played an important role in
generating fluctuations in aggregate real
variables and explaining the correlation
between real and nominal variables (for
example, Lucas, 1972). More recently, business
cycle research has been focused on a class of
models in which fluctuations associated with
the business cycle are the equilibrium outcome
of competitive economies that are subject to
exogenous technology shocks. In these real
business cycle models, as originally developed
by Finn Kydland and Edward Prescott (1982)
and John Long and Charles Plosser (1983), there
is a complete set of contingent claims markets
and money does not enter. Considering the
importance attributed to money in earlier
neoclassical and monetarist business cycle
theories, it is perhaps surprising that these real
models have been able to claim so much success
in replicating the characteristics of aggregate
data while abstracting from a role for money.
This does not imply that money is unimportant
for the evolution of real economic variables,
733
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE INFLATION TAX
734
THE AMERICAN ECONOMIC REVIEW
but it is true that the exact role for money in
these models is an open and somewhat
controversial question.
Not surprisingly, given that the correlation
between money and output is a time-honored
statistical regularity, the absence of money in
real business cycle models has been a source of
discomfort for many macroeconomists. One
reaction to this, for example, by Ben Bernanke
(1986) and Martin Eichenbaum and Kenneth
Singleton (1986) among others, has been to
reexamine the evidence that money “causes”
changes in output. Another approach has been
to construct models where money plays an
essentially passive role but in which the money
output correlation can be explained by
distinguishing different roles for money (for
example, inside and outside money) as in King
and Plosser (1984) and Jeremy Greenwood and
Gregory Huffman (1987). Yet another reaction
has been to argue that there is some role for
money over and above technology shocks. This
argument is pursued in Lucas (1987).
In this paper we study the quantitative
importance of money in a real business cycle
model where money is introduced in a way that
emphasizes the influence on real variables of
anticipated inflation operating through the inflation
tax. Money can have important real effects in
this setting: anticipated inflation will cause people
to substitute away from activities that require
cash, such as consumption, for activities that do
not require cash, such as leisure. Nevertheless,
this structure does not provide any role for
unanticipated money or “sticky price”
mechanisms, which many believe to be the most
important channel of influence of money on the
real economy. We analyze the consequence of
the distortion due to anticipated inflation for real
variables and estimate the magnitude of the
welfare losses that result.
In the following sections we describe,
calibrate, and simulate a simple one-sector
stochastic optimal growth model with a real
economy identical to that studied by Gary
Hansen (1985). The real time-series generated
by the model fluctuate in response to exogenous
technology shocks. The model inment lottery
201
SEPTEMBER 1989
that permits some agents to be unemployed. With
the latter features, the model implies a degree of
intertemporal substitution that is consistent with
observed fluctuations without contradicting
microeconomic evidence from panel studies. In
addition, the indivisible labor assumption is
consistent with the observation that most of the
fluctuation in aggregate hours worked is due to
fluctuations in employment rather than
fluctuations in the average hours worked of an
employed worker.
Money is introduced into the model using a
cash-in-advance constraint. Economies with this
feature have been studied extensively by Alan
Stockman (1981), Lucas (1982), Lucas and
Nancy Stokey (1983, 1987) and Lars Svensson
(1985). The cash-in-advance constraint applies
only to the consumption good. Leisure and
investment in our model are credit goods. Thus,
if agents in this economy wish to reduce cash
holdings in response to higher inflation, they
can only do so by reducing consumption.
In the next section of the paper we lay out
the details of our model and describe the
competitive equilibrium. Solving for an
equilibrium in this economy is more difficult
than in other real business cycle economies
because the inefficiency imposed by the cashin-advance constraint rules out the use of
invisible hand arguments based on the second
welfare theorem. In Section III we describe how
we solve for an equilibrium directly using a
method described in Kydland (1987).
In Section IV of the paper we present the
results of some simulations of the model under
various assumptions about the behavior of the
monetary growth rate. Our purpose here is to
use our model as an experimental device to
study the effect of certain parameter
interventions.1 We take a model whose statistical
properties have been studied previously and
examine how injections of money, operating
through a cash-in-advance constraint, alter the
conclusions derived from this purely real
1
See Thomas Cooley and Stephen LeRoy
(1985) for a discussion of parameter and variable interventions.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
202
SOME EXTENSIONS
VOL. 79 NO. 4
COOLEY AND HANSEN: THE INFLATION TAX
economy. In this model, when money is supplied
optimally, the real economy and its associated
steady-state paths and cyclical characteristics are
identical to those in Hansen (1985). This follows
from the fact that when money is supplied
optimally, the cash-in-advance constraint is not
binding. By varying the rate of growth of the
money supply we can study how the real
allocation and the comovements among variables
are altered. In addition we are able to measure
the welfare costs of the inflation tax.
The results of the experiments just described
are easily summarized. When money is
supplied according to a constant growth rate
rule that implies positive nominal interest rates,
individuals substitute leisure for goods, output
and investment fall, and the steady-state capital
stock is lower. The features of the business cycle
are unchanged by these constant growth rates.
We also report the results of experiments in
which money is supplied not at a constant rate
but erratically with characteristics that mimic
historical experience. In these simulations, the
cyclical behavior of real variables are altered
slightly: consumption becomes more variable
relative to income and the price level becomes
quite volatile. In addition, the correlations
between these variables and output become
smaller in absolute value. It is encouraging that
with these changes the cyclical properties of
the model more closely match U.S. postwar
experience.
Using definitions described in Section IV
we estimate the welfare cost due to the inflation
tax of a sustained moderate (10 percent)
inflation to be about 0.4 percent of GNP using
M1 as the relevant definition of money and a
quarter as the period over which it must be
held. This is very close to estimates that have
been suggested by others. We find the welfare
costs to be much lower, about 0.1, when the
relevant definition of money is the monetary
base and the period over which it is constrained
to be held is a month.
Perhaps the most striking implication of our
model for the steady-state behavior of economic
aggregates is that employment rates should be
lower in the long run in high inflation
economies. This possibility, stated somewhat
differently as the proposition that in the long
735
run the Phillips curve slopes upward, has been
suggested by others, most notably by Friedman
(1977). We present evidence that, for a cross
section of developed economies during the
period 1976–1985, average inflation rates and
average employment rates are negatively
correlated.
The conclusions drawn from our simulations
reflect only the costs and consequences of money
that are due to the inflation tax: there are no
informational problems created by the money
supply process. We conclude that if money does
have a major effect on the cyclical properties of
the real economy it must be through channels
that we have not explored here.
I. A Cash-in-Advance Model with Production
The economy studied is a version of the indivisible labor model of Hansen (1985) with
money introduced via a cash-in-advance constraint applied to consumption. That is, consumption is a “cash good” while leisure and
investment are “credit goods,” in the terminology of Lucas and Stokey (1983, 1987). In this
section we describe the economy and define a
competitive equilibrium. In the next section
we describe how an equilibrium can be computed using a linear-quadratic approximation
of the economy.
We assume a continuum of identical
households with preferences given by the utility
function,
(1)
where ct is consumption and is leisure in time
t. Households are assumed to be endowed with
one unit of time each period and supply labor to
a firm which produces the goods. Households
are also engaged in accumulating capital which
they rent to the firm.
We assume that households enter period t
with nominal money balances equal to mt-1 that
are carried over from the previous period. In
addition, these balances are augmented with a
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE INFLATION TAX
736
THE AMERICAN ECONOMIC REVIEW
lump-sum transfer equal to (gt-1)Mt-1, where Mt
is the per capita money supply in period t. The
money stock follows a law of motion
(2)
In this paper, we study two economies. In the
first, the gross growth rate of money, gt, is assumed to be constant. In the other economy, the
log of the gross growth rate of the money supply
evolves according to an autoregression of the form:
(3)
In equation (3), ξt is an iid random variable with
and variance
where
mean
is the unconditional mean of logarithm of
the growth rate gt. It is assumed that gt is revealed
to all agents at the beginning of period t.
Households are required to use these
previously acquired money balances to purchase
the nonstorable consumption good. That is, a
household’s consumption choice must satisfy the
constraint,
203
SEPTEMBER 1989
nonconvex. However, following Richard
Rogerson (1988), we convexify the economy by
assuming that agents trade employment lotteries.
That is, households sell contracts which specify a
probability of working in a given period, πt, rather
than selling their labor directly. Since all agents
are identical, they will all choose the same πt.
Thus, a fraction πt of the households will work h0
hours and the remaining (1-πt) households will
be unemployed during period t. A lottery
determines which of the households work and
which do not. Thus, per capita hours worked in
period t is given by
(5)
The market structure described above implies
that the period utility function of the representative household as a function of consumption
and hours worked is given by5
(4)
where pt is the price level at time t. In this paper,
attention is focused on examples where this constraint always holds with equality. A sufficient
condition for this constraint to be binding is that
the gross growth rate of money, gt, always exceeds the discount factor, β. Our examples will
satisfy this condition.2 In our view this assumption is not unreasonable given the observed behavior of the actual money supply.3
As in Hansen (1985), labor is assumed to be
indivisible. This means that households can work
some given positive number of hours, h0<1, or
not at all. They are not allowed to work an
intermediate number of hours.4 Under usual
market interpretations, this assumption implies
that the consumption set of households is
2
It can be shown from the first-order conditions of the household’s problem that the cashin-advance constraint will be binding (the
Lagrange multiplier associated with constraint
(3) will be positive) if and only if Et(1/gt+1)<1/β.
This condition follows from the use of log utility
and the timing assumptions.
3
In addition, to relax this assumption
would considerably complicate our solution
procedure, forcing us to consider the possibility of both corner and interior solutions.
4
The indivisible labor assumption implies that
all changes in total hours worked are due to
changes in the number of workers. Although over
half of the variance in total hours in the United
States is unambiguously due to fluctuations in
employment, there is still a significant percentage
that is due to fluctuation in average hours. A
model that allows for adjustment along both of
these margins is studied in J.O.Cho and Cooley
(1988).
5
This derivation makes use of the fact that
consumption is the same whether or not the household is employed. This result, which holds in equilibrium, follows from the separability of (1) in consumption and leisure and is shown formally in
Hansen (1985). It is possible to have unemployed
agents consume less than employed without significantly affecting the results obtained from the
model by assuming a nonseparable utility function (see Hansen, 1986). A more robust feature of
this model is that utility is higher for unemployed
individuals than for employed. Rogerson and
Randall Wright (1988) show that this implication
can be reversed if leisure is assumed to be an inferior good. It is unclear how one would reverse this
implication without significantly affecting the other
results obtained from the model.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
204
SOME EXTENSIONS
VOL. 79 NO. 4
COOLEY AND HANSEN: THE INFLATION TAX
We rewrite this as,
(6)
737
an exogenous shock to technology that follows a
law of motion given by
(10)
where
In the remainder of this section, we will discuss
the problem faced by a representative agent with
preferences given by (6) as a stand-in for the
individual household with preferences given by
(1) who is subject to the labor indivisibility restriction.
This representative household must choose
consumption, investment (xt), and nominal
money holdings subject to the following budget
constraint:6
(7)
where et is an iid random variable with mean 0
We assume that zt, like gt, is
and variance
revealed to all agents at the beginning of period t.
The firm seeks to maximize profit, which is
The first-order
equal to
conditions for the firm’s problem yield the
following functions for the wage rate and rental
rate of capital:
(11)
(12)
In this equation, wt and rt are the wage rate and
rental rate of capital, respectively. Investment is
undertaken to augment the capital stock (kt)
owned by the household. The capital stock
obeys the following law of motion:
(8)
The firm in our economy produces output
(Yt) using the constant returns to scale technology:
(9)
Capital letters are used to distinguish per capita
variables that a competitive household takes as
parametric from individual-specific variables that
are chosen by the household.7 The variable zt is
6
This budget constraint incorporates the fact
that consumption and investment sell at the same
price even though one is a cash good and the other
a credit good. This is because, from the point of
view of the seller, sales of both credit goods and
cash goods result in cash that will be available for
spending at the same time in the following period.
Although cash good sales in a given period result
in cash receipts in the same period, this cash can
not be spent until the next period.
A change in variables is introduced so that the
problem solved by the households will be stationary. That is, let
In addition, let
be the equilibrium
maximized present value of the utility stream of
the representative household who enters the period with a fraction of per capita money balances equal to and a capital stock equal to k
when the aggregate state is described by z, g, and
K. Implicit in the functional form of V are the
equilibrium aggregate decision rules (H and X)
as functions of the
and the pricing function
aggregate state, which is taken as given by the
households. The function V must satisfy
Bellman’s equation (primes denote next period
values)8
(13)
7
In equilibrium these will be the same.
Note that the solution to the firm’s profit
maximization problem has been substituted into
this problem through the functions w( ) and r( ).
8
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE INFLATION TAX
738
THE AMERICAN ECONOMIC REVIEW
subject to
205
SEPTEMBER 1989
compute an equilibrium for our cash-in-advance
economy.9
Kydland’s method involves computing a
linear-quadratic approximation to the
household’s problem (13). This dynamic
programming problem is then solved by
iterating on Bellman’s equation, requiring that
the second equilibrium condition (refer to the
above definition of equilibrium) hold at each
step of this recursive procedure. In the
remainder of this section, we outline in more
detail how this procedure is implemented in
our particular case.
The first step is to substitute the nonlinear
constraints, (14) and (15), into the household’s
utility function (6). This is done by first
eliminating c by substituting (15) into (14) and
(6). The resulting budget constraint is
(14)
(15)
(16)
(17)
(18)
(19)
and c, x,
nonnegative and 0≤h≤1. In addition,
X, H, and are given functions of (z, g, K).
A stationary competitive equilibrium for this
economy consists of a set of decision rules, c(s), x(s),
and h(s) (where
a set of aggregate decision rules, X(S) and H(S)
and a
(where S=(z, g, K)), a pricing function
value function V(s) such that:
(i) the functions V, X, H, and satisfy (13)
and h are the associated set of decision
and c, x,
rules;
when k=K and
(ii) x=X, h=H, and
and
(iii) the functions c(s) and x(s) satisfy
c(s)+x(s)=Y(S) for all s.
(20)
Because of the constant returns to scale technology, requiring that the functions w and r be of
the form (11) and (12) guarantees that equilibrium condition (iii) is satisfied.
The constraint (20) can be substituted into
the utility function (6) by eliminating h. However,
we must first eliminate H. This is done by
aggregating (20) and solving for H. Using (11)
and (12), this implies
(21)
II. Solution Method
In Hansen (1985) it was possible to compute an
equilibrium indirectly by solving for the (unique)
equal weight Pareto optimal allocation and invoking the second welfare theorem. In order to
obtain an analytic solution to the problem, a
linear-quadratic approximation to this nonlinear problem was formed, making it possible to
compute linear decision rules. Unfortunately, it
is not possible to invoke the second welfare theorem to compute an equilibrium for the economy
studied in this paper. This is because money
introduces a “wedge of inefficiency” (in the words
of Lucas, 1987) that forces one to solve for an
equilibrium directly. To get around this, we apply the method described in Kydland (1987) to
Equation (21) can be substituted into (20), and
the result substituted into (6). The return func-
9
This method is similar to the method of
Kydland and Prescott (1977), which is described
in some detail in Thomas Sargent (1981) and
Charles Whiteman (1983). In addition to
Kydland’s method, a number of other approaches to solving dynamic equilibrium models with distortions have been recently proposed
in the literature. Examples include papers by
David Bizer and Kenneth Judd (1988), Marianne
Baxter (1988), and Wilbur Coleman (1988).
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
206
SOME EXTENSIONS
VOL. 79 NO. 4
COOLEY AND HANSEN: THE INFLATION TAX
tion for the household’s dynamic programming
problem is now given by the following expression:
(22)
739
relating U to S. We start with a guess for the
matrix V, call it V 0 , and consider the
maximization problem on the right side of (23).
Once the laws of motion, (16) through (19),
have been substituted into the objective, we
obtain from the first-order condition for u the
linear decision rule
(24)
In order to obtain an analytic solution to this
problem, the above nonlinear return function
(22) is approximated by a quadratic function in
the neighborhood of the steady state of the certainty problem. This approximation technique
is described in detail in Kydland and Prescott
(1982). The state vector of the resulting linearquadratic dynamic programming problem is
K, k)T and the individuals′ decision
s=(1, z, g,
In addition,
(or control) vector is
also
the economywide variables
enter the quadratic return function. Thus, after
computing the quadratic approximation of (22),
Bellman’s equation for the household’s problem (13) become10
(23)
By imposing the equilibrium conditions, x=X,
and k=K; we can obtain, from (24),
a linear expression for U in terms of S that we
take as our candidate. That is, we obtain
(25)
To compute the value function for the next iteration, we evaluate the objective function on the
right side of (23) using our initial guess V0, the
function relating U to S (25) and the household’s
decision rule (24).11 This provides a quadratic
form, sTV1s, that is used as the value function for
the next iteration. This procedure is repeated
until Vj+1 is sufficiently close to Vj to claim that
the iterations have converged.
Once this process has converged, we obtain
the following equilibrium expressions for X and
( is equal to the inverse of consumption in
an equilibrium where the cash-in-advance
constraint is always binding):
(26)
(27)
subject to (16)–(19) and a linear function that
describes the relationship between U and
S=(1, z, g, K) T perceived by the agents in the
model.
To solve for an equilibrium, we iterate on
this quadratic version of Bellman’s equation.
This procedure must involve choosing a
candidate for the perceived linear function
10
This form for Bellman’s equation incorporates both certainty equivalence and the fact
that the value function will be quadratic.
Examples of these decision rules for particular
parameterizations of the money supply rule are
11
For the parameterizations studied in this paper it is not always possible to invert the firstorder conditions to obtain an expression like (24).
However, it is always possible to obtain equation
(25). Therefore, when evaluating (23), we used
(25) and, in place of (24) the equilibrium expressions for the components of u(
and x=X).
The first-order conditions are satisfied given the
way in which (25) is constructed and the fact that
the coefficients on k and
always turn out to
equal zero in these first-order conditions.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE INFLATION TAX
740
THE AMERICAN ECONOMIC REVIEW
given in the Appendix. These equations, which
determine investment and consumption, along
with the laws of motion (16) through (18), the
expression for hours worked (21), and the technology (9), are used to simulate artificial timeseries for various parameterizations of the gt process. These experiments are discussed in the next
section.
III. Results
We use the artificial economy just described to
study the interaction between money and the
real sector of the economy. We first describe the
cyclical behavior of our economy under various money supply rules. We then use the model
to measure the welfare costs of anticipated inflation. Finally, we look for confirmation of the
implied steady-state behavior of high and low
inflation economies in cross-section data on several developed countries.
A. Cyclical Properties
Statistics summarizing the cyclical behavior of
our model economy under various money supply rules, as well as statistics summarizing the
cyclical behavior of actual U.S. time-series, are
presented in Table 1. We will begin by describing how these statistics are computed and then
proceed to interpret our results.
The first panel of Table 1 shows the (percent)
standard deviations of the set of endogenous
variables and their correlations with output that
characterize recent U.S. quarterly data. These
provide some basis for comparison with the
results of our experiments although we wish to
stress that ours is not a data matching exercise
but an experimental simulation of a model
economy. We use quarterly data from 1955,3
to 198 4,1 on real G N P, consumption,
investment, capital stock, hours worked,
productivity, and two measures of the price
level, the CPI and GNP deflator. 12 Before
12
The series for real GNP, investment, hours
worked, and the price level were taken from the
Citibase database. The hours series is based on
207
SEPTEMBER 1989
computing statistics, the data (both actual and
simulated) are logged and detrended using the
Hodrick-Prescott filter. The use of this
detrending procedure enables us to maintain
comparability with prior real business cycle
studies by Kydland and Prescott (1982) and
Hansen (1985).
In order to derive results from the artificial
economies, we follow Kydland and Prescott
(1982) by choosing parameter values based on
growth observations and the results of studies
using microeconomic data. In order to make
comparisions with Hansen (1985) meaningful,
we set the parameters describing preferences and
technology to the same values used in that study.
Those values, which were chosen under the
assumption that the length of a period is one
quarter, are β=0.99, θ=0.36, δ=0.025, B=2.86,
and γ=0.95. The standard deviation of ε, σε, is
set equal to 0.00721 so that the standard
deviation of the simulated output series is close
to the standard deviation of the actual output
series. We experiment with different values for
the parameters describing the money supply
process.
Given a set of parameter values, simulated
time-series with 115 observations (the number
of observations in the data sample) are computed
using the method described in the previous
section. These series are then logged and filtered
and summary statistics calculated. We simulate
the economy 50 times and the averages of the
statistics over these simulations are reported. In
addition, we report the sample standard
deviations of these statistics, which are given in
parentheses.
The columns of the second panel of Table 1
show the percent standard deviations and
correlations that result from all of the simulations
information from the Current Population Survey.
Productivity is output divided by hours worked.
The data on the capital stock include government capital stock and private capital stock (housing) as well as producers’ durables and structures. The consumption series includes nondurables and services plus an imputed flow of services from the stock of durables. The consumption and capital stock series were provided by
Larry Christiano.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
208
SOME EXTENSIONS
VOL. 79 NO. 4
COOLEY AND HANSEN: THE INFLATION TAX
741
TABLE 1—STANDARD DEVIATIONS IN PERCENT AND CORRELATIONS WITH
OUTPUT FOR U.S. AND ARTIFICIAL ECONOMICS
a
The U.S. time-series reported on are real GNP, consumption of nondurables and services, plus
the flow of services from durables, gross private domestic investment (all in 1982 dollars). The
capital stock series includes nonresidential equipment and structures, residential structures, and
government capital. The hours series is total hours for persons at work in nonagricultural industries
as derived from the Current Population Survey. Productivity is output divided by hours. All series are
seasonally adjusted, logged, and detrended. The output, investment, hours, and price-level series
were taken from the Citibase database. The consumption and capital stock series were provided by
Larry Christiano.
b
The percent standard deviations and correlations with output are sample means of statistics
computed for each of 50 simulations. Each simulation is 115 periods long, which is the same
number of periods as the U.S. sample. The sample standard deviations of these statistics are in
parentheses. Each simulated time-series was logged and detrended using the same procedure applied to the U.S. sample before the statistics were calculated.
of our model economy where the money supply
grows at a constant rate. These results confirm
that when money is supplied at a constant growth
rate, even one that implies a high average
inflation rate, the features of the business cycle
are unaffected. In particular, the statistics
summarizing the behavior of the real variables
are the same as would be obtained in the same
model without money—the “indivisible labor”
model of Hansen (1985).
The remaining two panels of Table 1 show
the results of simulations with an erratic money
supply. That is, we assume a money supply rule
of the form (3). We calibrate this money supply
process (that is, choose values for α and σζ) so
that the money supply varies in a way that is
broadly consistent with postwar experience. We
proceed by assuming that the Fed draws money
growth rates from an urn with the draws being
serially correlated, as in equation (3). We
determined the characteristics of that urn from
data on M1 and the regression (standard errors
in parentheses)
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE INFLATION TAX
742
THE AMERICAN ECONOMIC REVIEW
where M1 is the average quarterly value. We
intentionally averaged to smooth the data somewhat and increase the implied persistence.13 The
results of this regression lead us to set a equal to
0.48 and σξ equal to 0.009. To ensure that the
gross rate of money growth always exceeds the
discount factor, as is required for the cash-inadvance constraint to be always binding, we draw
ζt from a lognormal distribution. This implies
that log(gt) will never become negative.
The statistics reported in Table 1 show that
volatility of the money supply has a small but
significant impact on the cyclical characteristics
of the economy. Virtually all the effect of
volatility in the money supply is in the standard
deviations of consumption and prices and their
correlation with output. In particular,
consumption and prices become more volatile
and their correlation with output becomes
smaller in absolute value. It is worth noting that
the numbers in these panels are more in keeping
with historical experience (see first panel) than
are the results from constant growth rate
economies. In addition, comparing the third and
fourth panels we find that, although the price
level does become more volatile, increases in
the average growth rate of money has little effect
on the cyclical properties of the real variables.
B. Welfare Costs of the Inflation Tax
In this section estimates of the welfare costs of
the inflation tax are presented that are derived
by comparing steady states of our growth model
assuming different growth rates of the money
supply.14 Measuring the welfare costs of antici13
This equation is open to criticism as a description of the historical sample. Although we
cannot reject its adequacy, there may be a leftover moving average piece in the residuals. This
in turn could imply that some portion of the
innovation in the money growth rate is permanent. See, for example, G.William Schwert
(1987). We chose to ignore this because the estimated autoregression seems to capture the features that are appropriate for our experiment.
14
A somewhat similar approach to that taken
here appears in a recent paper by Jean Pierre
Danthine, John Donaldson, and Lance Smith
209
SEPTEMBER 1989
pated inflation is an old issue in macroeconomics. Martin Bailey (1956) provided a classic answer to this question by considering the area
under the demand curve for money, the welfare
triangle, evaluated at an interest rate embodying
the steady-state rate of inflation as a measure of
the net loss to individuals from the inflation tax.
Stanley Fischer (1981) and Robert Lucas (1981)
updated Bailey’s estimates and they supply a
thoughtful discussion of some of the awkward
assumptions underlying the welfare triangle approach (for example, that government expenditures are financed by non-distorting taxes). They
also discuss some of the subsidiary costs of inflation that are ignored by those calculations.
We chose to measure the welfare costs by
comparing steady states because, as explained
above, the cyclical characteristics of this
economy are unaffected by the average growth
rate of the money stock. Thus, our discussion of
welfare is based on the steady-state properties of
a version of our economy where the money
supply grows at a constant rate and the
technology shock in equation (9) is replaced by
its unconditional mean.
The welfare costs for various annual inflation
rates, along with the associated steady-state values
for output, consumption, investment, the capital
stock, and hours worked, are presented in Table
2. We show results based on two different
assumptions on the length of time that the cashin-advance constraint is binding. The numbers
displayed in the top panel reflect the assumption
that the relevant period over which individuals
are constrained to hold money is a quarter. This
is consistent with the calibration of the model in
the previous section. In addition, if we assume a
unitary velocity as is implied by our model and
if we assume that the “cash good” corresponds
to consumption of nondurables and services then
(1987). Their model differs from ours in that
money appears directly in the utility function
and they do not include labor in their model. In
addition, they assume that capital depreciates
fully each period. They also demonstrate a decline in welfare with inflation, but do so using
simulations of their economy rather than comparing steady states.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
210
SOME EXTENSIONS
VOL. 79 NO. 4
COOLEY AND HANSEN: THE INFLATION TAX
743
TABLE 2—STEADY STATES AND WELFARE COSTS ASSOCIATED
WITH VARIOUS ANNUAL GROWTH RATES OF MONEY
this would be consistent with defining money as
M1, based on evidence from the 1980s.15
The results given in the bottom panel of Table
2 are based on the assumption that the relevant
period over which individuals are constrained
to hold money is a month. It turns out that
monthly consumption of nondurables and
services corresponds roughly to the monetary
base during the 1980s. The steady states in this
second panel were computed using different
parameter values for the discount factor and
depreciation rate of capital in order to maintain
comparability to the quarterly results. The values
assigned were β=0.997 and δ=0.008, which are
the monthly rates that correspond to the
quarterly rates assumed above. We also scale
the production function to reflect monthly
output levels by multiplying the right-hand side
15
This conclusion is based on the fact that
the ratio of the stock of M1 to quarterly consumption of nondurables and services has been
close to one since the late 1970s. Unfortunately,
this result does not hold over a long period of
time-the ratio has been as high as 3 early in the
postwar period. The same caveat applies to the
observation concerning the monetary base made
below.
of equation (9) by 1/3. The values for the gross
growth rate of the money supply (g) that
correspond to the desired annual inflation rates
are also different for the monthly model. We
indicate these values in the table.
The welfare measure we use is based on the
increase in consumption that an individual
would require to be as well off as under the
Pareto optimal allocation. The Pareto optimal
allocation for our economy is equivalent to the
equilibrium allocation for the same economy
without the cash-in-advance constraint, or,
equivalently, for a version of the model where
the money supply grows at a rate such that the
cash-in-advance constraint is never binding. It
turns out that for the model studied in this paper,
the cash-in-advance constraint is not binding if
the gross growth rate of money is equal to the
discount factor, β.16 To obtain a measure of the
16
We restrict the growth rate, g, to be greater
than or equal to β. This ensures that nominal
interest rates will not be negative (see Lucas and
Stokey, 1987). When we set g=β, the initial price
level is no longer uniquely determined. However, the real allocation and rate of inflation are
uniquely determined and the allocation is Pareto
optimal.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE INFLATION TAX
744
THE AMERICAN ECONOMIC REVIEW
welfare loss associated with growth rates that
are larger than β , we solve for ΔC in the
equation
(28)
where
is the level of utility attained (in the
steady state) under the Pareto optimal allocation
(g=β), and C* and H* are the steady-state consumption and hours associated with the growth
rate in question (some g>β).
The results of the welfare calculations
expressed as a percent of steady-state real
output (ΔC/Y) and steady-state real
consumption (ΔC/C) are shown in the bottom
rows of both panels of Table 2. The welfare
cost of a moderate (10 percent) inflation is
0.387 percent of GNP when the period over
which individuals are constrained is a quarter.
This magnitude may be compared to the
estimates of 0.3 percent provided by Stanley
Fischer or 0.45 percent obtained by Robert
Lucas based on an approximation of the area
under a money demand function. 17 It is
interesting that their exercise, which holds
output constant but allows velocity to vary,
yields the same answer as our exercise which
holds velocity constant but allows output to
vary. While an estimate of roughly 0.4 percent
of GNP sounds small, at current levels of GNP
it would amount to $15.2 billion of real GNP.
The welfare costs of very high inflation rates,
which are not uncommon throughout the
world, seem extremely high.
If the relevant period over which
individuals are constrained is a month then
the welfare costs are considerably reduced
being only 0.11 percent at a 10 percent annual
inflation rate and slightly more than 1.5 percent
at a 400 percent annual inflation rate.
Evidently the period over which individuals
are constrained, and by implication the
17
Fischer and Lucas use different definitions
of money (high-powered money and M1, respectively) and different estimates of the interest
elasticity.
211
SEPTEMBER 1989
definition of the money balances on which
individuals are taxed, make a big difference
in the welfare costs of inflation.
Since there is a big difference in the estimates
it is worth considering what some of the biases
might be. Our larger estimates come from
assuming that individuals are constrained for
one quarter, which is roughly consistent with
assuming that the appropriate monetary
aggregate is M1. However, a large part of M1
consists of checkable deposits. To the extent that
these earn competitive interest they will be
shielded from the inflation tax. At the other
extreme, the monetary base consists of currency
and reserves. Since these are clearly subject to
the inflation tax, the monthly data provides a
lower bound on the magnitude of the welfare
loss. It seems reasonable that in economies with
sustained high inflations many individuals will
be able to shield themselves against the inflation
tax. If the institutions did not exist to facilitate
this, one would expect them to evolve in very
high inflation economies. For this reason, our
model may not be very reliable for analyzing
hyperinflation. On the other hand these estimates
abstract from many of the subsidiary costs of
inflation that are believed to be important.
Among these are distortions caused by
nonneutralities in the tax system and adjustment
costs or confusion caused by the variability of
inflation.
C. Steady-State Implications of Inflation
As shown in Table 2, anticipated inflation has a
significant influence on the steady-state path of
the economy. Steady-state consumption, output,
hours, investment, and the capital stock are all
lower whenever the growth rate of the money
supply exceeds the optimal level (g=β). The
consumption of leisure increased because agents
substitute this “credit good” for the consumption good in the face of a positive inflation tax
on the latter. Lower hours worked leads to lower
output and therefore lower consumption, investment, and capital stock. The share of output
allocated to investment does not change with
higher inflation. This result is obtained despite
the fact that consumption is a cash good and
investment is a credit good since, in the steady
state, investment will provide consumption in
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
212
SOME EXTENSIONS
VOL. 79 NO. 4
COOLEY AND HANSEN: THE INFLATION TAX
745
of employment rates on inflation rates. There is
a statistically significant negative correlation
between inflation rates and employment rates.
The coefficient of the inflation rate in a regression
of the employment rate on the inflation rate and
a constant is -0.5 with a standard error of 0.17.
The most extreme observation in the graph
pertains to Chile. When that is eliminated the
conclusions are essentially unchanged; the
coefficient of inflation is -0.44 with a standard
error 0.22. These results suggest that the
phenomenon displayed in our model economy
may not be counterfactual.
IV. Conclusions
FIGURE 1. AVERAGE EMPLOYMENT
AND INFLATION RATES, 1976–1985
the future that will be subject to exactly the same
inflation tax as consumption today.
A striking implication of higher inflation rates
in our model economy is that they are
accompanied by lower employment rates.18 The
“menu of choices” available to the monetary
authority involves giving up low inflation only
to obtain higher unemployment. This result, that
the operational Phillips curve is upward sloping,
is also obtained by Greenwood and Huffman
(1987) for their model economy. Friedman
(1977) in his Nobel lecture presented some
evidence for this phenomenon by plotting data
from several countries. Here we present some
statistical evidence that supports the negative
correlation between employment rates and
inflation rates using a cross section of countries.
Figure 1 shows the relation between the
average rate of employment and the average
rate of inflation from 1976 to 1985 for 23
countries.19 The solid line depicts the regression
18
The variable HOURS in Table 2, which
corresponds to per capita hours worked, is actually the employment rate multiplied by a constant
(h0), given the assumption of indivisible labor.
19
The countries are Austria, Belgium, Denmark, Finland, France, W.Germany, Greece, Ire-
In this paper we incorporate an interesting paradigm for money holding, the cash-in-advance
model, in a stochastic optimal growth model
with an endogenous labor leisure decision. We
have shown that the solution and simulation of
such a model is quite tractable. The model and
solution procedure provide a basis for studying
the influence of inflation on the path of the real
economy and its cyclical characteristics. In addition, the solution procedure we have used
could be employed to study the effects of other
distortions as well.
We have used this model as the basis for
estimating the welfare cost of the inflation tax
and studying the long-run features of economies
with different inflation rates. The fact that our
estimates are well within the range of estimates
obtained by other methods and that the
empirical implications are confirmed in crosssectional data is very encouraging. This suggests
to us that the approximations and simplifications
we have made in writing down a tractable model
of a competitive economy incorporating money
may not be too serious. This is not to argue that
econometric estimation of many of the
parameters we have simply specified might not
land, Italy, Netherlands, Norway, Portugal, Spain,
Sweden, Switzerland, U K, Canada, United
States, Australia, New Zealand, Japan, Chile, and
Venezuela. Population data are taken from Summers and Allan Heston (1988) and the remainder of the data are taken from the International
Labor Office (1987).
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE INFLATION TAX
746
THE AMERICAN ECONOMIC REVIEW
yield further insights into these problems. What
we find appealing about this approach is that all
the features of the economy, from the decision
rules to the specification of technology and
preferences are explicit. Nothing is hidden. This
makes it a valuable environment for
experimental exercises like those considered
here, and for positive exercises, for example
where one would model the behavior of the
monetary authority.
Although we have shown that anticipated
inflation can have significant effects on the longrun values of real variables, our model economy
predicts that the business cycle will be the same
in a high inflation economy as in a low inflation
economy. When money is supplied erratically,
the characteristics of the business cycle are altered
somewhat. These changes in the characteristics
of the cycle occur solely because of changes in
allocations that result from the changing
conditional expectation of inflation. Unexpected
inflation has no role in this model. However, we
speculate that the most important influence of
money on short-run fluctuations are likely to
stem from the influence of the money supply
process on expectations of relative prices, as in
the natural rate literature. That is, if money does
have a significant effect on the characteristics of
the cycle it is likely to come about because the
behavior of the monetary authority has serious
informational consequences for private agents.
APPENDIX
Decision Rules for Selected Cases
Constant Growth Rate
213
SEPTEMBER 1989
Autoregressive Growth Rate
REFERENCES
Bailey, Martin J., “The Welfare Cost of Inflationary Finance,” Journal of Political Economy,
April 1956, 64, 93–110.
Baxter, Marianne, “Approximating Suboptimal
Dynamic Equilibria: A Euler Equation
Approach,” reproduced, University of
Rochester, 1988.
Bernanke, Ben S., “Alternative Explanations of
the Money-Income Correlation,” CarnegieRochester Conference on Public Policy,
K.Brunner and A.Meltzer, eds., Autumn
1986, 25, 49–99.
Bizer, David and Judd, Kenneth, “Capital Accumulation, Risk, and Uncertain Taxation,”
reproduced, University of Chicago, 1988.
Cho, Jang-Ok and Cooley, Thomas F., “Employment and Hours Over the Business
Cycle,” reproduced, University of Rochester, 1988.
Citibank, “Citibase Economic Database,” New
York: Citibank, 1978.
Coleman, Wilbur, “Money, Interest, and Capital in a Cash-in-Advance Economy,”
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
214
SOME EXTENSIONS
VOL. 79 NO. 4
COOLEY AND HANSEN: THE INFLATION TAX
reproduced, Board of Governors of the Federal Reserve System, 1988.
Cooley, Thomas F. and LeRoy, Stephen F.,
“Atheoretical Macroeconomics: A Critique,” Journal of Monetary Economics, November 1985, 16, 283–308.
Danthine, Jean Pierce, Donaldson, John and
Smith, Lance, “On the Superneutrality
of Money in a Stochastic Dynamic Macroeconomic Model,” Journal of Monetary
Economics, July 1987, 20, 475–500.
Eichenbaum, Martin and Singleton, Kenneth,
“Do Equilibrium Real Business Cycle
Theories Explain Post-War U.S. Business
Cycles?” in Macroeconomics Annual, Stanley
Fischer, ed., Vol. I, 1986, Cambridge: MIT
Press.
Fischer, Stanley, “Towards an Understanding
of The Costs of Inflation,” Carnegie-Rochester Conference on Public Policy, K. Brunner
and A.Meltzer, eds. Autumn 1981, 15, 5–
42.
Friedman, Milton, “Nobel Lecture: Inflation
and Unemployment,” Journal of Political
Economy, June 1977, 85, 451–72.
____, ____, and Heller, Walter W., Monetary
vs. Fiscal Policy, New York: Norton, 1969.
____,____, and Schwartz, Anna J., A Monetary
History of the United States, 1967–1960,
Princeton: Princeton University Press,
1963.
Greenwood, Jeremy and Huffman, Gregory,
“A Dynamic Equilibrium Model of Inflation and Unemployment,” Journal of Monetary Economics, March 1987, 19, 203–28.
Hansen, Gary D., “Indivisible Labor and the
Business Cycle,” Journal of Monetary Economics, November 1985, 16, 309–28.
____, “Growth and Fluctuations,” reproduced,
UCLA, 1986.
International Labour Office, Yearbook of Labour
Statistics, 1987, Geneva.
King, Robert and Plosser, Charles, “Money,
Credit, and Prices in a Real Business Cycle
Economy,” American Economic Review, June
1984, 74, 363–80.
Kydland, Finn E., “The Role of Money in a
Competitive Model of Business Fluctuations,” reproduced, Carnegie-Mellon University, 1987.
747
____and Prescott, Edward C., “Rules Rather
Than Discretion: The Inconsistency of Optimal Plans,” Journal of Political Economy, June
1977, 85, 473–91.
___ and ____, “Time to Build and Aggregate
Fluctuations,” Econometrica, November
1982, 50, 1345–70.
Long, John and Plosser, Charles, “Real Business Cycles,” Journal of Political Economy,
February 1983, 91, 39–69.
Lucas, Robert E., Jr., “Expectations and the
Neutrality of Money,” Journal of Economic
Theory, April 1972, 4, 103–24.
____, “Discussion of Towards an Understanding of the Costs of Inflation: II.” CarnegieRochester Conference on Public Policy, K.Brunner
and A.Meltzer, eds., Autumn 1981, 15, 43–
52.
____, “Interest Rates and Currency Prices in a
Two-Country World,” Journal of Monetary Economics, November 1982, 12, 335–60.
_____, Models of Business Cycles, New York: Basil
Blackwell, 1987.
____ and Stokey, Nancy L., “Optimal Fiscal
and Monetary Policy in an Economy Without Capital,” Journal of Monetary Economics,
July 1983, 12, 55–93.
____, “Money and Interest in a Cash-in-Advance
Economy,” Econometrica, May 1987, 55,
491–514.
Prescott, Edward C., “Can The Cycle Be Reconciled with a Consistent Theory of Expectations?,” reproduced, Federal Reserve
Bank of Minneapolis, 1983.
____, “Theory Ahead of Business Cycle Measurement,” Federal Reserve Bank of Minneapolis Quarterly Review, Fall 1986, 10, 9–22.
Rogerson, Richard, “Indivisible Labor, Lotteries and Equilibrium,” Journal of Monetary
Economics, January 1988, 21, 3–16.
____ and Wright, Randall, “Involuntary Unemployment in Economies with Efficient
Risk Sharing,” Journal of Monetary Economics,
November 1988, 22, 501–15.
Sargent, Thomas, “Lecture Notes on Filtering, Control, and Rational Expectations,”
reproduced, University of Minnesota,
1981.
Schwert, G.William, “The Effects of Model
Specification on Some Tests for Unit
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE INFLATION TAX
748
THE AMERICAN ECONOMIC REVIEW
Roots,” Journal of Monetary Economics, July
1987, 20, 73–103.
Stockman, Alan C., “Anticipated Inflation
and the Capital Stock in a Cash-in-Advance Economy,” Journal of Monetary Economics, November 1981, 8, 387–93.
Summers, Robert and Heston, Alan, “A New
Set of International Comparisons of Real
Product and Prices for 130 Countries,
215
SEPTEMBER 1989
1950–1985,” Review of Income and Wealth,
March 1988, 34, Supplement, 1–25.
Svensson, Lars E.O., “Money and Asset Prices
in a Cash-in-Advance Economy,” Journal
of Political Economy, October 1985, 93, 919–
44.
Whiteman, Charles, Linear Rational Expectations
Models: A User’s Guide, Minneapolis: University of Minnesota Press, 1983.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
Part IV
The methodology of
equilibrium business cycle
models
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
CHAPTER 12
219
Scand. J. of Economics 93(2), 161–178, 1991
The Econometrics of the General Equilibrium
Approach to Business Cycles*
Finn E.Kydland
Carnegie-Mellon University, Pittsburgh PA, USA
Edward C.Prescott
Federal Reserve Bank and University of Minnesota, Minneapolis MN, USA
Abstract
The founding fathers of the Econometric Society defined econometrics to be quantitative
economic theory. A vision of theirs was the use of econometrics to provide quantitative
answers to business cycle questions. The realization of this dream required a number of
advances in pure theory—in particular, the development of modern general equilibrium
theory. The econometric problem is how to use these tools along with measurement to
answer business cycle questions. In this essay, we review this econometric development and
contrast it with the econometric approach that preceded it.
I. Introduction
Early in this century American institutionists and members of the German
historical school attacked—and rightfully so—neoclassical economic theory
for not being quantitative. This deficiency bothered Ragnar Frisch and
motivated him, along with Irving Fisher, Joseph Schumpeter, and others,
to organize the Econometric Society in 1930. The aim of the society was
to foster the development of quantitative economic theory—that is, the
development of what Frisch labeled econometrics. Soon after its inception,
the society started the journal Econometrica. Frisch was the journal’s first
editor and served in this capacity for 25 years.
In his editorial statement introducing the first issue of Econometrica (1933),
Frisch makes it clear that his motivation for starting the Econometric
* We acknowledge useful comments of Javier Diaz-Giménez on an early draft. This
research was partly supported by a National Science Foundation Grant. The views
expressed herein are those of the authors and not necessarily those of the Federal
Reserve Bank of Minneapolis or the Federal Reserve System.
Scand. J. of Economics 1991
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
220
METHODOLOGY
162 F.E.Kydland and E.C.Prescott
Society was the “unification of theoretical and factual studies in economics”
(p. 1). This unification of statistics, economic theory, and mathematics,
he argues, is what is powerful. Frisch points to the bewildering mass of
statistical data becoming available at that time, and asserts that in order
not to get lost “we need the guidance and help of a powerful theoretical
framework. Without this no significant interpretation and coordination
of our observations will be possible” (ibid., p. 2).
Frisch speaks eloquently about the interaction between theory and
observation when he says “theory, in formulating its abstract quantitative
notions, must be inspired to a larger extent by the technique of observation.
And fresh statistical and other factual studies must be the healthy element
of disturbance that constantly threatens and disquiets the theorist and
prevents him from coming to rest on some inherited, obsolete set of
assumptions” (ibid.). Frisch goes on to say that
this mutual penetration of quantitative economic theory and statistical
observation is the essence of econometrics. (ibid., p. 2).
To summarize the Frisch view, then, econometrics is quantitative
neoclassical theory with a basis in facts.
Forty years after founding the Econometric Society, Frisch (1970)
reviewed the state of econometrics. In this review he discusses what he
considers to be “econometric analysis of the genuine kind” (p. 163), and
gives four examples of such analysis. None of these examples involve the
estimation and statistical testing of some model. None involve an attempt
to discover some true relationship. All use a model, which is an abstraction
of a complex reality, to address some clear-cut question or issue.
It is interesting to note that, in his 1933 editorial statement, Frisch
announced that each year Econometrica would publish four surveys of “the
significant developments within the main fields that are of interest to the
econometrician” (ibid., p. 3). These fields are general economic theory
(including pure economics), business cycle theory, statistical technique,
and, finally, statistical information. We find it surprising that business
cycle theory was included in this list of main fields of interest to
econometricians. Business cycles were apparently phenomena of great
interest to the founders of the Econometric Society.
Frisch’s (1933) famous, pioneering work, which appears in the Cassel
volume, applies the econometric approach he favors to the study of business
cycles. In this paper, he makes a clear distinction between sources of
shocks on the one hand, and the propagation of shocks on the other. The
main propagation mechanism he proposes is capital-starting and carry-on
activities in capital construction, both of them features of the production
technology. Frisch considers the implications for duration and amplitude
of the cycles in a model that he calibrates using available micro data to
Scand. J. of Economics 1991
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE GENERAL EQUILIBRIUM APPROACH
The econometrics of the general equilibrium approach
221
163
select the numerical values for the parameters. Making the production
technology with capital accumulation a central element of the theory has
its parallel in modern growth theory.
There are many other papers dating from the 1930s that study business
cycle models. In these papers, however, and in those of the 1940s and
1950s, little progress was made beyond what Frisch had already done.
The main reason was that essential theoretical tools, in particular ArrowDebreu general equilibrium theory, statistical decision theory, modern
capital theory, and recursive methods had yet to be developed. The modern
electronic computers needed to compute the equilibrium processes of
dynamic stochastic model economies were also unavailable. Only after
these developments took place could Frisch’s vision be carried out.
In this paper, we review the development of econometric business cycle
theory, with particular emphasis on the general equilibrium approach
(which was developed later). Crucial to this development was the systematic
reporting of national income and product accounts, along with time series
of aggregate inputs and outputs of the business sector. Section II is a
review of this important development in factual studies. In Section III we
review what we call the system-of-equations approach to business cycle
theory. With this approach, a theory of the business cycle is a system of
dynamic equations which have been measured using the tools of statistics.
Section IV is a review of the general equilibrium approach to business
cycle theory. General equilibrium models have people or agents who have
preferences and technologies, and who use some allocation mechanism.
The crucial difference between the general equilibrium and the systemof-equations approaches is that which is assumed invariant and about
which we organize our empirical knowledge. With the system-of-equations
approach, it is behavioral equations which are invariant and are measured.
With the general equilibrium approach, on the other hand, it is the
willingness and ability of people to substitute that is measured. In Section
V we illustrate the application of this econometric approach to addressing
specific quantitative questions in the study of business cycles. Section VI
contains some concluding comments.
II. National Income and Product Accounts
An important development in economics is the Kuznets-Lindahl-Stone
national income and product accounts. Together with measures of
aggregate inputs to the business sector, these accounts are the aggregate
time series that virtually define the field of macroeconomics—which we
see as concerned with both growth and business cycle fluctuations. The
Kuznets-Lindahl-Stone accounting system is well-matched to the general
Scand. J. of Economics
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
1991
222
METHODOLOGY
164 F.E.Kydland and E.C.Prescott
equilibrium framework because there are both household and business
sectors, with measures of factor inputs to the business sector and of goods
produced by the business sector, as well as measures of factor incomes and
expenditures on products.
An examination of these time series reveals some interesting regularities—
in particular, a number of ratios which remain more or less constant.
These growth facts led Robert Solow to develop a neoclassical growth
model which simply and elegantly rationalized these facts. Solow’s
structure was not fully neoclassical, however, because the consumptionsavings decision was behaviorally determined rather than being the result
of maximizing behavior subject to constraints. With the consumptionsavings decision endogenized, Solow’s growth model does become fully
neoclassical, with agents’ maximizing subject to constraints and market
clearing. This structure can be used to generate time series of national
income and product accounts.
Aggregate data present other features that are of interest to economists,
such as the more volatile movements in the time series. During the 1950s and
1960s, neoclassical theory had not evolved enough to allow economists to
construct computable general equilibrium models with fluctuations. Lacking
the necessary tools, economists adopted an empirical approach and searched
for laws of motion governing these variables. They hoped this research
procedure would result in empirically determined laws which would
subsequently be rationalized within the neoclassical paradigm. In the natural
sciences, for example, empirically determined laws have often subsequently
been rationalized at a deeper theoretical level, and it was hoped that this
would also be the case in macroeconomics. In the following section we briefly
review the econometrics of this approach to business cycle fluctuations.
III. The System-of-Equations Approach
Tjalling Koopmans, who was influenced by Frisch and might even be
considered one of his students, argued forcefully in the late 1940s for
what he called the econometric approach to business cycle fluctuations.
At the time, it was the only econometric approach. The general equilibrium
approach to the study of business cycles had yet to be developed. But
since the approach Koopmans advocated is no longer the only one, another
name is needed for it. As it is the equations which are invariant and
measured, we label this approach the system-of-equations approach.1
1
Koopmans subsequently became disillusioned with the system-of-equations approach. When asked in the late 1970s by graduate students at the University of
Minnesota in what direction macroeconomics should go, Koopmans is reported by
Zvi Eckstein to have said they should use the growth model.
Scand. J. of Economics 1991
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE GENERAL EQUILIBRIUM APPROACH
The econometrics of the general equilibrium approach
223
165
In the 1930s, there were a number of business cycle models or theories.
These logically complete theories were a dynamic set of difference equations
that could be used to generate time series of the aggregate variables of
interest. Notable examples include Frisch’s (1933) model in Cassel’s
volume, Tinbergen’s (1935) suggestions on quantitative business cycles,
and Samuelson’s (1939) multiplier-accelerator model. One problem with
this class of models is that the quantitative behavior of the model depended
upon the values of the coefficients of the variables included in the equations.
As Haberler (1949) points out in his comment on Koopmans’ (1949) paper,
the stock of cyclical models (theories) is embarrassingly large. Give any
sophomore “a couple of lags and initial conditions and he will construct
systems which display regular, damped or explosive oscillation…as desired”
(p. 85). Pure theory was not providing sufficient discipline, and so it is
not surprising that Koopmans advocated the use of the statistics discipline
to develop a theory of business fluctuations.
System-of-Equations Models
As Koopmans (1949, p. 64) points out, the main features of the system-ofequations models are the following: First, they serve as theoretical exercises
and experiments. Second, the variables involved are broad aggregates,
such as total consumption, the capital stock, the price level, etc. Third,
the models are “logically complete, i.e., they consist of a number of
equations equal to the number of variables whose course over time is to be
explained”. Fourth, the models are dynamic, with equations determining
the current values of variables depending not only on current values of
other variables but also on the values of beginning-of-period capital stocks
and on lagged variables. Fifth, the models contain, at most, four kinds of
equations, which Koopmans calls structural equations. The first type of
equations are identities. They are valid by virtue of the definition of the
variables involved. The second type of equations are institutional rules, such
as tax schedules. The third type are binding technology constraints, that is,
production functions. The final type are what Koopmans calls behavioral
equations, which represent the response of groups of individuals or firms to
a common economic environment. Examples are a consumption function,
an investment equation, a wage equation, a money demand function, etc.
A model within this framework is a system-of-equations. Another
requirement, in addition to the one that the number of variables equal the
number of equations, is that the system have a unique solution. A final
requirement is that all the identities implied by the accounting system for
the variables in the model hold for the solution to the equation system;
that is, the solution must imply a consistent set of national income and
product accounts.
Scand. J. of Economics 1991
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
224
METHODOLOGY
166 F.E.Kydland and E.C.Prescott
Statistical Measurement of Equations
The behavior of these models depends crucially on the numerical magnitudes
of the coefficients of the variables and of the time lags. This leads to attempts
to estimate these parameters using time series of the variables being modeled.
Given that the estimation of these coefficients is a statistical exercise, a
probability model is an additional completeness requirement. For that
purpose, a residual random disturbance vector typically is added, with one
component for each behavioral equation. For statistical completeness, the
probability distribution of this disturbance vector must be specified up to
some set of parameters. Only then can statistical methods be applied to
estimating the coefficients of the behavioral equations and the parameters of
the disturbance distribution. The crucial point is that the equations of the
macroeconometric model are the organizing principle of the system-ofequations approach. What is measured is the value of the coefficients of the
equations. The criterion guiding the selection of the values of the coefficients
is essentially the ability of the resulting system of equations to mimic the
historical time series. The issue of which set of equations to estimate is
settled in a similar fashion. The criterion guiding the selection of equations
is in large part how well a particular set can mimic the historical data.
Indeed, in the 1960s a student of business cycle fluctuations was successful
if his particular behavioral equation improved the fit of, and therefore
replaced, a currently established equation.
The Rise and the Fall of the System-of-Equations Approach
With the emergence of a consensus on the structure of the system of
equations that best described the behavior of the aggregate economy, the
approach advocated by Koopmans became totally dominant in the 1960s.
This is well-illustrated by the following statement of Solow’s, quoted by
Brunner (1989, p. 197):
I think that most economists feel that the short run macroeconomic
theory is pretty well in hand… The basic outlines of the dominant
theory have not changed in years. All that is left is the trivial job of
filling in the empty boxes [the parameters to be estimated] and that will
not take more than 50 years of concentrated effort at a maximum.
The reign of this system-of-equations macroeconomic approach was not
long. One reason for its demise was the spectacular predictive failure of
the approach. As Lucas and Sargent (1978) point out, in 1969 these models
predicted high unemployment would be associated with low inflation.
Counter to this prediction, the 1970s saw a combination of both high
unemployment and high inflation. Another reason for the demise of this
Scand. J. of Economics 1991
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE GENERAL EQUILIBRIUM APPROACH
The econometrics of the general equilibrium approach
225
167
approach was the general recognition that policy-invariant behavioral
equations are inconsistent with the maximization postulate in dynamic
settings. The principal reason for the abandonment of the system-ofequations approach, however, was advances in neoclassical theory that
permitted the application of the paradigm in dynamic stochastic settings.
Once the neoclassical tools needed for modeling business cycle fluctuations
existed, their application to this problem and their ultimate domination
over any other method was inevitable.
IV. The General Equilibrium Approach
A powerful theoretical framework was developed in the 1950s and 1960s
that built upon advances in general equilibrium theory, statistical decision
theory, capital theory, and recursive methods. Statistical decision theory
provided a logically consistent framework for maximization in a dynamic
stochastic environment. This is what was needed to extend neoclassical
theory, with its maximization assumption, to such environments. Another
crucial development was the extension of general equilibrium theory to
dynamic stochastic models, with the simple yet important insight that
commodities could be indexed not only by type, but also by date and event.
This important insight was made by Arrow and Debreu (1954), who had
important precursors in the work of Hicks (1939) and, particularly, in that
of Lindahl (1929)—who had previously effectively extended competitive
theory to dynamic environments. Subsequently, recursive methods, with
their Markovian structure, were developed. These methods simplified the
use of this dynamic framework and, in particular, its extension to stochastic
general equilibrium analyses; see, for example, Stokey and Lucas (1989).
Perhaps just as important as the development of tools for carrying out
aggregate equilibrium analysis was the access to better and more systematic
national income and product accounts data. In his review of growth theory,
Solow (1970) lists the key growth facts which guided his research in growth
theory in the 1950s. These growth facts were the relative constancy of
investment and consumption shares of output, the relative constancy of
labor and capital income shares, the continual growth of the real wage
and output per capita, and the lack of trend in the return on capital.
Solow (1956), in a seminal contribution, developed a simple model
economy that accounted for these facts. The key to this early theory was
the neoclassical production function, which is a part of the general
equilibrium language. Afterwards the focus of attention shifted to
preferences, with the important realization that the outcome of the CassKoopmans optimal growth model could be interpreted as the equilibrium
of a competitive economy in which the typical consumer maximizes utility
and the markets for both factors and products clear at every date.
Scand. J. of Economics 1991
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
226
METHODOLOGY
168 F.E.Kydland and E.C.Prescott
General Equilibrium Models
By general equilibrium we mean a framework in which there is an explicit
and consistent account of the household sector as well as the business sector.
To answer some research questions, one must also include a sector for the
government, which is subject to its own budget constraint. A model within
this framework is specified in terms of the parameters that characterise
preferences, technology, information structure, and institutional
arrangements. It is these parameters that must be measured, and not some
set of equations. The general equilibrium language has come to dominate
in business cycle theory, as it did earlier in public finance, international
trade, and growth. This framework is well-designed for providing quantitative
answers to questions of interest to the business cycle student.
One of these important questions, which has occupied business cycle
theorists since the time of Frisch and Slutzky, is how to determine which
sources of shocks give rise to cycles of the magnitudes we observe. To provide
reliable answers to this and similar questions, abstractions are needed that
describe the ability and willingness of agents to substitute commodities, both
intertemporally and intratemporally, and within which one can bring to bear
statistical or factual information. One of these abstractions is the neoclassical
growth model. This model has proven useful in accounting for secular facts.
To understand business cycles, we rely on the same ability and willingness of
agents to substitute commodities as those used to explain the growth facts.
We are now better able than Frisch was more than 50 years ago to calibrate
the parameters of aggregate production technology. The wealth of studies on
the growth model have shown us the way. To account for growth facts, it may
be legitimate to abstract from the time allocation between market and nonmarket
activities. To account for business cycle facts, however, the time allocation is
crucial. Thus, good measures of the parameters of household technology are
needed if applied business cycle theory is to provide reliable answers.
The Econometrics of the General Equilibrium Approach
The econometrics of the general equilibrium approach was first developed
to analyze static or steady-state deterministic models. Pioneers of this
approach are Johansen (1960) and Harberger (1962). This framework was
greatly advanced by Shoven and Whalley (1972), who built on the work of
Scarf (1973). Development was impeded by the requirement that there be a
set of excess-demand functions, which are solved to find the equilibrium
allocations. This necessitated that preference and technology structures have
very special forms for which closed-form supply and demand functions
existed. Perhaps these researchers were still under the influence of the systemof-equations approach and thought a model had to be a system of supply
Scand. J. of Economics 1991
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE GENERAL EQUILIBRIUM APPROACH
The econometrics of the general equilibrium approach
227
169
and demand functions. These researchers lacked the time series needed
to estimate these equations. Given that they could not estimate the
equations, they calibrated their model economy so that its static equilibrium
reproduced the sectoral national income and product accounts for a base
year. In their calibration, they used estimates of the elasticity parameters
obtained in other studies.
Their approach is ill-suited for the general equilibrium modeling of
business fluctuations because dynamics and uncertainty are crucial to
any model that attempts to study business cycles. To apply general
equilibrium methods to the quantitative study of business cycle
fluctuations, we need methods to compute the equilibrium processes of
dynamic stochastic economies, and specific methods for the stochastic
growth model economy. Recursive competitive theory and the use of
linear-quadratic economies are methods that have proven particularly useful.
These tools make it possible to compute the equilibrium stochastic processes
of a rich class of model economies. The econometric problem arises in the
selection of the model economies to be studied. Without some restrictions,
virtually any linear stochastic process on the variables can be rationalized
as the equilibrium behavior of some model economy in this class. The
key econometric problem is to use statistical observations to select the
parameters for an experimental economy. Once these parameters have
been selected, the central part of the econometrics of the general equilibrium
approach to business cycles is the computational experiment. This is the
vehicle by which theory is made quantitative. The experiments should be
carried out within a sensible or appropriate model economy that is capable
of addressing the question whose answer is being sought. The main steps
in econometric analyses are as follows: defining the question; setting up
the model; calibrating the model; and reporting the findings.
Question
To begin with, the research question must be clearly defined. For example,
in some of our own research we have looked at quantifying the
contribution of changes in a technology parameter, also called Solow
residuals, as a source of U.S. postwar business cycles. But we refined it
further. The precise question asked is how much variation in aggregate
economic activity would have remained if technology shocks were the
only source of variation. We emphasize that an econometric, that is,
quantitative theoretic analysis, can be judged only relative to its ability to
address a clear-cut question. This is a common shortcoming of economic
modeling. When the question is not made sufficiently clear, the model
economy is often criticized for being ill-suited to answer a question it was
never designed to answer.
Scand. J. of Economics 1991
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
228
170
METHODOLOGY
F.E.Kydland and E.C.Prescott
Model Economy
To address a specific question one typically needs a suitable model
economy for addressing the specified question. In addition to having a
clear bearing on the question, tractability and computability are essential
in determining whether the model is suitable. Model-economy selection
depends on the question being asked. Model-economy selection should
not depend on the answer provided. Searching within some parametric
class of economies for the one that best fits some set of aggregate time
series makes little sense. Unlike the system-of-equations approach, no
attempt is made to determine the true model. All model economies are
abstractions and are by definition false.
Calibration
The model has to be calibrated. The necessary information can sometimes
be obtained from data on individuals or households. An example of such
information is the average fraction of discretionary time household
members who are, or who potentially are, labor market participants actually
spent in market activity. In many other cases, the required information
easily can be obtained from aggregate nonbusiness-cycle information. The
task often involves merely computing some simple averages, such as growth
relations between aggregates. This is the case for inventory-output and
capital-output ratios, and long-run fractions of the various G NP
components to total output, among others.
In some cases, history has provided sufficiently dramatic price
experiments which can be used to determine, with a great deal of
confidence, an elasticity of substitution. In the case of labor and capital as
inputs in the aggregate business production function, and also in the case
of consumption and leisure as inputs to household production, the large
real-wage increase over several decades in relation to the prices of the
other input, combined with knowledge about what has happened to the
expenditure shares on the respective inputs, provides this kind of
information. Because the language used in these business cycle models is
the same as that used in other areas of applied economics, the values of
common parameters should be identical across these areas and typically
have been measured by researchers working in these other areas. One can
argue that the econometrics of business cycles described here need not be
restricted to general equilibrium models. In fact it is in the stage of
calibration where the power of the general equilibrium approach shows
up most forcefully. The insistence upon internal consistency implies that
parsimoniously parameterized models of the household and business sector
display rich dynamic behavior through the intertemporal substitution
arising from capital accumulations and from other sources.
Scand. J. of Economics 1991
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE GENERAL EQUILIBRIUM APPROACH
The econometrics of the general equilibrium approach
229
171
Computational Experiments
Once the model is calibrated, the next step is to carry out a set of
computational experiments. If all the parameters can be calibrated with a
great deal of accuracy, then only a few experiments are needed. In practice,
however, a number of experiments are typically required in order to provide
a sense of the degree of confidence in the answer to the question. It often
happens that the answer to the research question is robust to sizable
variations in some set of parameters and conclusions are sharp, even though
there may be a great degree of uncertainty in those parameters. At other
times, however, this is not the case, and without better measurement of
the parameters involved, theory can only restrict the quantitative answer
to a large interval.
Findings
The final step is to report the findings. This report should include a
quantitative assessment of the precision with which the question has been
answered. For the question mentioned above, the answer is a numerical
estimate of the fraction of output variability that would have remained if
variations in the growth of the Solow residual were the only source of
aggregate fluctuation. The numerical answer to the research question, of
course, is model dependent. The issue of how confident we are in the
econometric answer is a subtle one which cannot be resolved by computing
some measure of how well the model economy mimics historical data.
The degree of confidence in the answer depends on the confidence that
is placed in the economic theory being used.
V. Two Applications to Business Cycle Theory
We illustrate the econometrics of the general equilibrium approach to
business cycle theory with two examples. The first example, credited to
Lucas (1987) and Imrohoroglu (1989), addresses the question of
quantifying the costs of business cycle fluctuations. An important feature
of the quantitative general equilibrium approach is that it allows for explicit
quantitative welfare statements, something which was generally not
possible with the system-of-equations approach that preceded it. The second
example investigates the question of how large business cycle fluctuations
would have been if technology shocks were the only source of fluctuations.
This question is also important from a policy point of view. If these shocks
are quantitatively important, an implication of theory is that an important
component of business cycle fluctuations is a good, not a bad.
Scand. J. of Economics 1991
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
230
METHODOLOGY
172 F.E.Kydland and E.C.Prescott
Costs of Business Cycle Fluctuations
The economy Lucas uses for his quantitative evaluation is very simple.
There is a representative or stand-in household and a random endowment
process of the single consumption good. The utility function is standard,
namely, the expected discounted value of a constant relative risk aversion
utility function. Equilibrium behavior is simply to consume the
endowment. Lucas determines how much consumption the agent is willing
to forgo each period in return for the elimination of all fluctuations in
consumption. Even with an extreme curvature parameter of 10, he finds
that when the endowment process is calibrated to the U.S. consumption
behavior, the cost per person of business cycle fluctuations is less than
one-tenth of a per cent of per-capita consumption.
This model abstracts from important features of reality. There is no
investment good, and consequently no technology to transform the date
t consumption good into the date t+1 consumption good. As the costs of
fluctuation are a function of the variability in consumption and not in
investment, abstracting from capital accumulation is appropriate relative
to the research question asked. What matters for this evaluation is the
nature of the equilibrium consumption process. Any representative-agent
economy calibrated to this process will give the same answer to the question,
so it makes sense to deal with the simplest economy whose equilibrium
consumption process is the desired one. This is what Lucas does.
Introducing the time-allocation decision between market and nonmarket
activities would change the estimate, since the agent would have the
opportunity to substitute between consumption and leisure. The
introduction of these substitution opportunities would result in a reduction
in the estimated cost of business cycle fluctuations as leisure moves
countercyclically. But, given the small magnitude of the cost of business
cycle fluctuations, even in a world without this substitution opportunity,
and given that the introduction of this feature reduces the estimate of this
cost, there is no need for its inclusion.
In representative-agent economies, all agents are subject to the same
fluctuations in consumption. If there is heterogeneity and all idiosyncratic
risk is allocated efficiently, the results for the representative and
heterogeneous agent economies coincide. This would not be the case if
agents were to smooth consumption through the holding of liquid assets as
is the case in the permanent income theory. Imrohoroglu (1989) examines
whether the estimated costs of business cycle fluctuations are significantly
increased if, as is in fact the case, people vary their holdings of liquid assets
in order to smooth their stream of consumption. She modifies the Lucas
economy by introducing heterogeneity and by giving each agent access to a
technology that allows that agent to transform date t consumption into date
t+1 consumption. Given that real interest rates were near zero in the
Scand. J. of Economics 1991
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE GENERAL EQUILIBRIUM APPROACH
The econometrics of the general equilibrium approach
231
173
fifty-odd years from 1933 to 1988, the nature of the storage technology
chosen is that one unit of the good today can be transferred into one unit
of the good tomorrow. She calibrates the processes on individual
endowments to the per-capita consumption process, to the variability of
annual income across individuals, and to the average holdings of the
liquid asset—also across individuals. For her calibrated model economy,
she finds the cost of business cycles is approximately three times as large
as that obtained in worlds with perfect insurance of idiosyncratic risk.
But three times a small number is still a small number.
Technology Shocks as Source of Fluctuations
One source of shocks suggested as far back as in Wicksell (1907) is
fluctuations in technological growth. In the 1960s and 1970s, this source
was dismissed by many as being unlikely to play much of a role in the
aggregate. Most researchers accepted that there could be considerable
variation in productivity at the industry level, but they believed that
industry-level shocks would average out in the aggregate. During the
1980s, however, this source of shocks became the subject of renewed
interest as a major source of fluctuations, in large part supported by
quantitative economic theory. The question addressed, then, was how
much would the U.S. postwar economy have fluctuated if technological
shocks were the only source of aggregate fluctuations?
Our selection of a model economy to address this question follows.
First we extended the neoclassical growth model to include leisure as an
argument of the stand-in household’s utility function. Given that more
than half of business cycle fluctuations are accounted for by variations in
the labor input, introducing this element is crucial. Next we calibrated
the deterministic version of the model so that its consumption-investment
shares, factor income shares, capital output ratios, leisure-market time
shares, and depreciation shares matched the average values for the U.S.
economy in the postwar period. Throughout this analysis, constant
elasticity structures were used. As uncertainty is crucial to the question,
computational considerations led us to select a linear-quadratic economy
whose average behavior is the same as the calibrated deterministic constant
elasticity of substitution economy.
We abstracted from public finance considerations and consolidated the
public and private sectors. We introduced Frisch’s (1933) assumption of
time-to-build new productive capital. The construction period considered
was four periods, with new capital becoming productive only upon
completion, but with resources being used up throughout its construction.
Given the high volatility of inventory investment, inventory stocks were
included as a factor of production. We found, using the variance of Solow
residuals estimated by Prescott (1986), that the model economy’s output
Scand. J. of Economics 1991
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
232
174
METHODOLOGY
F.E.Kydland and E.C.Prescott
variance is 55 per cent as large as the corresponding variance for the U.S.
economy in the postwar period.
In the early 1980s, there was much discussion in the profession about
the degree of aggregate intertemporal substitution of leisure. The feeling
was that this elasticity had to be quite high in order for a market-clearing
model to account for the highly volatile and procyclical movements in
hours. This discussion may have started with the famous paper by Lucas
and Rapping (1969). Realizing that the standard utility function implied
a rather small elasticity of substitution, they suggested the possibility that
past leisure choices may directly affect current utility. Being sympathetic
to that view, we considered also a non-time-separable utility function as a
tractable way of introducing this feature. When lags on leisure are
considered, the estimate of how volatile the economy would have been if
technology shocks were the only disturbance increases from 55 to near 70
per cent. But, until there is more empirical support for this alternative
preference structure, we think estimates obtained using the economy with
a time-separable utility function are better. Unlike the system-of-equations
approach, the model economy which better fits the data is not the one
used. Rather, currently established theory dictates which one is used.
Probably the most questionable assumption of this theory, given the
question addressed, is that of homogeneous workers, with the additional
implication that all variation in hours occurs in the form of changes in
hours per worker. According to aggregate data for the U.S. economy,
only about one-third of the quarterly fluctuations in hours are of this
form, while the remaining two-thirds arise from changes in the number
of workers; see Kydland and Prescott (1989, Table 1).
This observation led Hansen (1985) to introduce the Rogerson (1988)
labor indivisibility construct into a business cycle model. In the Hansen
world all fluctuations in hours are in the form of employment variation.
To deal with the apparent nonconvexity arising from the assumption of
indivisible labor, the problem is made convex by assuming that the
commodity points are contracts in which every agent is paid the same
amount whether that agent works or not, and a lottery randomly chooses
who in fact works in every period. Hansen finds that with this labor
indivisibility his model economy fluctuates as much as did the U.S.
economy. Our view is that, with the extreme assumption of only
fluctuations in employment, Hansen overestimates the amount of aggregate
fluctuations accounted for by Solow residuals in the same way as our
equally extreme assumption of only fluctuations in hours per worker lead
us to an underestimation.
In Kydland and Prescott (1989), the major improvement on the 1982
version of the model economy is to permit variation both in the number of
workers and in the number of hours per worker. The number of hours a
Scand. J. of Economics 1991
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE GENERAL EQUILIBRIUM APPROACH
The econometrics of the general equilibrium approach
233
175
plant is operated in any given period is endogenous. The model also
treats labor as a quasi-fixed input factor by assuming costs of moving
people into and out of the business sector. Thus, in this model there is
what we interpret to be labor hoarding.
Without the cost of moving workers in and out of the labor force, a
property of the equilibrium turns out to be that all the hours variation is
in the form of employment change and none in hours per worker. In that
respect, it is similar to Hansen’s (1985) model. For this economy with no
moving costs, the estimate is that Solow residuals account for about 90 per
cent of the aggregate output variance. For this economy with moving
costs, we calibrated so that the relative variations in hours per worker and
number of workers matched U.S. data. With this degree of labor hoarding,
the estimate of the fraction of the cycle accounted for by Solow residuals
is reduced to 70 per cent.
A widespread and misguided criticism of our econometric studies, for
example, McCallum (1989), is that the correlation between labor productivity
and the labor input is almost one for our model economy while it is
approximately zero for the U.S. postwar economy. If we had found that
technology shocks account for nearly all fluctuations and that other factors
were unimportant, the failure of the model economy to mimic the data in
this respect would cast serious doubt on our findings. But we did not find
that the Solow technology shocks are all-important. We estimate that these
technology shocks account for about 70 per cent of business cycle
fluctuations. If technology shocks account for 70 per cent, and some other
shocks which are orthogonal to technology shocks account for 30 per cent,
theory implies a correlation between labor productivity and the labor input
near zero. Christiano and Eichenbaum (1990) have established this formally
in the case that the other shock is variations in public consumption. But the
result holds for any shock that is orthogonal to the Solow technology shocks.
The fact that this correlation for our model economy and the actual data
differ in the way they do adds to our confidence in our findings.
The estimate of the contribution of technology shocks to aggregate shocks
has been found to be robust to several modifications in the model economy.
For example, Greenwood, Hercowitz, and Huffman (1988) permit the
utilization rate of capital to vary and to affect its depreciation rate, while all
technology change is embodied in new capital; Danthine and Donaldson
(1989) introduce an efficient-wage construct; Cooley and Hansen (1989)
consider a monetary economy with a cash-in-advance constraint; and RiosRull (1990) uses a model calibrated to life cycle earnings and consumption
patterns. King, Plosser, and Rebelo (1988) have non-zero growth. Gomme
and Greenwood (1990) have heterogenous agents with recursive preferences
and equilibrium risk allocations. Benhabib, Rogerson, and Wright (1990)
incorporate home production. Hornstein (1990) considers the implications
Scand. J. of Economics 1991
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
234
176
METHODOLOGY
F.E.Kydland and E.C.Prescott
of increasing returns and monopolistic competition. In none of these cases
is the estimate of the contribution of technology shocks to aggregate
fluctuations significantly altered.
VI. Concluding Remarks
Econometrics is by definition quantitative economic theory—that is,
economic analyses which provide quantitative answers to clear-cut
questions. The general equilibrium econometric methodology is centered
around computational experiments. These experiments provide answers
to the questions posed in the model economies whose equilibrium elements
have been computed. The model economy selected should quantitatively
capture people’s ability and willingness to substitute and the arrangements
employed which are relevant to the question. We base our quantitative
economic intuition on the outcome of these experiments.
The dramatic advances in econometric methodology over the last 25
years have made it possible to apply fully neoclassical econometrics to the
study of business cycles. Already there have been several surprising
findings. Contrary to what virtually everyone thought, including the
authors of this review, technology shocks were found to be an important
contributor to business cycle fluctuations in the U.S. postwar period.
Not all fluctuations are accounted for by technology shocks, and monetary
shocks are a leading candidate to account for a significant fraction of the
unaccounted-for aggregate fluctuations. The issue of how to incorporate
monetary and credit factors into the structure is still open, with different
avenues under exploration. When there is an established monetary theory,
we are sure that general equilibrium methods will be used econometrically
to evaluate alternative monetary and credit arrangements.
References
Arrow, Kenneth J. & Debreu, Gerard: Existence of an equilibrium for a
competitive economy. Econometrica 22(3), 265–90, 1954.
Benhabib, Jess, Rogerson, Richard & Wright, Randall: Homework in
Macroeconomics I: Basic Theory, mimeo, 1990.
Brunner, Karl: The disarray in macroeconomics. In Forrest Capie & Geoffrey
E.Wood (eds.), Monetary Economics in the 1980s, MacMillan Press, New
York, 1989.
Christiano, Lawrence J. & Eichenbaum, Martin: Current real business cycle
theories and aggregate labor market fluctuations. DP 24, Institute for
Empirical Macroeconomics, Federal Reserve Bank of Minneapolis and
University of Minnesota, 1990.
Cooley, Thomas F. & Hansen, Gary D.: The inflation tax in a real business
cycle model. American Economic Review 79(4), 733–48, 1989.
Scand. J. of Economics 1991
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE GENERAL EQUILIBRIUM APPROACH
The econometrics of the general equilibrium approach
235
177
Danthine, Jean-Pierre & Donaldson, John B.: Efficiency wages and the real business
cycle. Cahier 8803, Départment d’économétri et d’économie politique,
Université de Lausanne, 1988; forthcoming in European Economic Review.
Frisch, Ragnar: Propagation problems and impulse problems in dynamic
economics. In Economic Essays in Honor of Gustav Cassel, London, 1933.
Frisch, Ragnar: Econometrics in the world of today. In W.A.Eltis, M.F.G.Scott
& J.N. Wolfe (eds.), Induction, Growth and Trade: Essays in Honour of Sir Roy
Harrod, Clarendon Press, Oxford, 152–66, 1970.
Greenwood, Jeremy, Hercowitz, Zvi & Huffman, Gregory W.: Investment,
capacity utilization and the business cycle. American Economic Review 78,
402–18, June 1988.
Gomme, Paul & Greenwood, Jeremy: On the cyclical allocation of risk. WP
462, Research Department, Federal Reserve Bank of Minneapolis, 1990.
Harberger, Arnold C: The incidence of the corporation income tax. Journal of
Political Economy 70(3), 215–40, 1962.
Haberler, Gottfried: “Discussion” of the econometric approach to business
fluctuations by Tjalling C.Koopmans. American Economic Review 39, 84–8,
May 1949.
Hansen, Gary D.: Indivisible labor and the business cycle. Journal of Monetary
Economics 16(3), 309–27, 1985.
Hicks, John R.: Value and Capital: An Inquiry into Some Fundamental Principles of
Economic Theory. Clarendon Press, Oxford, 1939.
Hornstein, Andreas: Monopolistic competition, increasing returns to scale,
and the importance of productivity changes. WP, University of Western
Ontario, 1990.
Imrohoroglu, Ayse: Costs of business cycles with indivisibilities and liquidity
constraints. Journal of Political Economy 97, 1364–83, Dec. 1989.
Johansen, Leif: A Multi-sectoral Study of Economic Growth. North-Holland,
Amsterdam, 1960.
King, Robert G., Plosser, Charles I. & Rebelo, Sergio T.: Production, growth
and business cycles II: New directions. Journal of Monetary Economics 21,
309–41, March/May 1988.
Koopmans, Tjalling C.: The econometric approach to business fluctuations.
American Economic Review 39, 64–72, May 1949.
Kydland, Finn E. & Prescott, Edward C.: Time to build and aggregate
fluctuations. Econometrica 50, 1345–70, 1982.
Kydland, Finn E. & Prescott, Edward C.: Hours and employment variation in
business cycle theory. DP 17, Institute for Empirical Macroeconomics,
Federal Reserve Bank of Minneapolis and University of Minnesota, 1989;
forthcoming in Economic Theory.
Kydland, Finn E. & Prescott, Edward C.: Business cycles: Real facts and a
monetary myth. Federal Reserve Bank of Minneapolis Quarterly Review, 3–18,
Spring 1990.
Lindahl, Erik: Prisbildningsproblemets uppläggning från kapitalteoretisk
synpunkt. Ekonomisk Tidskrift 31, 31–81, 1929. Translated as: The place
of capital in the theory of price, in Studies in the Theory of Money and Capital,
Farrar and Reinhart, New York, 269–350, 1929.
Lucas, Robert E., Jr.: Models of Business Cycles. Yrjö Jahnsson Lectures, Basil
Blackwell, Oxford and New York, 1987.
Scand. J. of Economics 1991
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
236
178
METHODOLOGY
F.E.Kydland and E.C.Prescott
Lucas, Robert E., Jr. & Rapping, Leonard A.: Real wages, employment and
inflation. Journal of Political Economy 77, 721–54, Sept./Oct. 1969.
Lucas, Robert E., Jr. & Sargent, Thomas J.: After Keynesian macroeconomics.
In After the Phillips Curve: Persistence of High Inflation and High Unemployment,
Conference Series No. 19, Federal Reserve Bank of Boston, 49–72, 1978.
McCallum, Bennett T.: Real business cycle models. In R.J.Barro (ed.), Modern
Business Cycle Theories, Harvard University Press, Boston, 16–50, 1989.
Prescott, Edward C.: Theory ahead of business cycle measurement. CarnegieRochester Conference Series on Public Policy 25, 11–44, 1986.
Rios-Rull, Jose Victor: Life cycle economies and aggregate fluctuations.
Preliminary draft, Carnegie-Mellon University, June 1990.
Rogerson, Richard: Indivisible labor, lotteries and equilibrium. Journal of Monetary
Economics 21, 3–16, Jan. 1988.
Samuelson, Paul A.: Interaction between the multiplier analysis and the principle
of acceleration. Review of Economic and Statistics 29, 75–8, May 1939.
Scarf, Herbert (with the collaboration of T.Hansen): Computation of Economic
Equilibria. Yale University Press, New Haven, 1973.
Schumpeter, Joseph: The common sense of econometrics. Econometrica 1, 5–
12, Jan. 1933.
Shoven, John B. & Whalley, John: A general equilibrium calculation of the
effects of differential taxation of income from capital in the U.S. Journal of
Public Economics 1 (3/4), 281–321, 1972.
Solow, Robert M.: A contribution to the theory of economic growth. Quarterly
Journal of Economics 70 (1), 65–94, 1956.
Solow, Robert M.: Technical change and the aggregate production function.
Review of Economics and Statistics 39(3), 312–20, 1957.
Solow, Robert M.: Growth Theory: An Exposition. Radcliffe Lectures, Clarendon
Press, Oxford, 1970.
Stokey, Nancy & Lucas, Robert E., Jr., with Prescott, Edward C.: Recursive
Methods in Economic Dynamics. Harvard University Press, Cambridge, MA,
1989.
Tinbergen, Jan: Annual survey: Suggestions on quantitative business cycle
theory. Econometrica 3, 241–308, July 1935.
Wicksell, Knut: Krisernas gåta. Statsøkonomisk Tidsskrift 21, 255–70, 1907.
Scand. J. of Economics 1991
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
CHAPTER 13
237
Journal of Economic Perspectives—Volume 10, Number 1—Winter 1996—Pages 69–85
The Computational Experiment:
An Econometric Tool
Finn E.Kydland and Edward C.Prescott
I
n a computational experiment, the researcher starts by posing a welldefined quantitative question. Then the researcher uses both theory
and measurement to construct a model economy that is a computer
representation of a national economy. A model economy consists of
households, firms and often a government. The people in the model
economy make economic decisions that correspond to those of their
counterparts in the real world. Households, for example, make
consumption and savings decisions, and they decide how much to work
in the market. The researcher then calibrates the model economy so that
it mimics the world along a carefully specified set of dimensions. Finally,
the computer is used to run experiments that answer the question.1
Such experiments have become invaluable tools in quantitative aggregate
theory.2 They are being used, for example, to estimate the quantitative
effects of trade liberalization policies, measure the welfare consequences
of changes in the tax system and quantify the magnitude and nature of
business cycle fluctuations induced by different types of shocks. In this
paper, we review the use of the computational experiment in economics.
1
Lucas (1980), in his paper on methods and problems in business cycle theory, explains the
need for computational experiments in business cycle research.
2
Shoven and Whalley (1972) were the first to use what we call the computational experiment
in economics. The model economies that they used in their experiments are static and have
many industrial sectors.
ƒ Finn E.Kydland is Professor of Economics, Carnegie-Mellon University, Pittsburgh, Pennsylvania, and
Research Associate, Federal Reserve Bank of Cleveland, Cleveland, Ohio. Edward C.Prescott is Professor
of Economics, University of Minnesota, and Research Associate, Federal Reserve Bank of Minneapolis,
both in Minneapolis, Minnesota.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
238
METHODOLOGY
70 Journal of Economic Perspectives
One immediate question that arises is whether the computational
experiment should be regarded as an econometric tool (for example, Gregory
and Smith, 1993). In the modern (narrow) sense of the term it is not, since it
isn’t used in the “measurement of economic relations” (Marschak, 1948, p. 1).
Yet it is an econometric tool in the original (broad) sense of the term (which
we prefer), since computational experiments are used to derive the
quantitative implications of economic theory (Frisch, 1933a, p. 1). In Kydland
and Prescott (1991a), we develop the position that the computational
experiment is an econometric tool, but here, we avoid this largely semantic
debate. Instead, we will simply state that the task of deriving the quantitative
implications of theory differs from that of measuring economic parameters.
Computational experiments are not unique to economic science—they are
heavily used in the physical sciences as well. In one crucial respect, however,
they do differ across the two disciplines. Unlike theory in the physical
sciences, theory in economics does not provide a law of motion. Rather,
economic theory provides a specification of people’s ability and willingness to
substitute among commodities. Consequently, computational experiments in
economics include the additional step of computing the equilibrium process in
which all of the model’s people behave in a way that is in each person’s best
interest—that is, economists must compute the equilibrium law of motion or
process of the model economy. Given the process governing the system, both
economic and physical science use the computer to generate realizations of
this process.
If the model is deterministic, only one possible equilibrium realization
exists for the path of the model economy. If the model economy has
aggregate uncertainty—as it must, for example, if the phenomena of interest
are business cycle fluctuations—then the model will imply a process governing
the random evolution of the economy. In the case of uncertainty, the
computer can generate any number of independent realizations of the
equilibrium stochastic process, and these realizations, along with statistical
estimation theory, are then used to measure the sampling distribution of any
desired set of statistics of the model economy.
Steps in an Economic Computational Experiment
Any economic computational experiment involves five major steps: pose a
question; use a well-tested theory; construct a model economy; calibrate the model
economy; and run the experiment. We will discuss each of these steps in turn.
Pose a Question
The purpose of a computational experiment is to derive a quantitative answer
to some well-posed question. Thus, the first step in carrying out a computational
experiment is to pose such a question. Some of these questions are concerned
with policy evaluation issues. These questions typically ask about the welfare
and distributive consequences of some policy under consideration. Other
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE COMPUTATIONAL EXPERIMENT
Finn E.Kydland and Edward C.Prescott
239
71
Table 1
Examples of Well-Posed Questions in Studies Using the Computational Experiment
questions are concerned with the testing and development of theory. These
questions typically ask about the quantitative implications of theory for some
phenomena. If the answer to these questions is that the predictions of theory
match the observations, theory has passed that particular test. If the answer is
that there is a discrepancy, a deviation from theory has been documented.
Still, other experiments are concerned with the sensitivity of previous findings
to the introduction of some feature of reality from which previous studies
have abstracted. Table 1 offers some examples of computational experiments
that seek to answer each of these types of questions. That table highlights the
fact that judging whether a model economy is a “good” abstraction can be
done only relative to the question posed.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
240
METHODOLOGY
72 Journal of Economic Perspectives
Use Well-Tested Theory
With a particular question in mind, a researcher needs some strong theory to
carry out a computational experiment: that is, a researcher needs a theory
that has been tested through use and found to provide reliable answers to a
class of questions. Here, by theory we do not mean a set of assertions about
the actual economy. Rather, following Lucas (1980), economic theory is defined
to be “an explicit set of instructions for building…a mechanical imitation system”
to answer a question. If the question is quantitative in nature, a computer
representation of the imitation system or economy is used, and extensive
computations are required to answer the posed question.
As one example, the computational experiments often carried out in
modern business cycle theory build upon the neoclassical growth framework.
Central to neoclassical growth theory is its use of an aggregate production
function, with the output of goods resulting from inputs of labor and capital.3
This framework has served well when dealing with growth within reasonably
stable economic institutions. With an explicit description of the household
sector, including its focus on the time-allocation decision, the neoclassical
growth model becomes an internally consistent framework for addressing
business cycle questions, as well as other questions of interest to
macroeconomists. The theory implies that when a model economy is
confronted with technology, public finance and terms-of-trade shocks, it
should display business cycle fluctuations of a quantitative nature similar to
those actually observed. In other words, modern business cycle models are
stochastic versions of neoclassical growth theory. And the fact that business
cycle models do produce normal-looking fluctuations adds dramatically to our
confidence in the neoclassical growth theory model—including the answers it
provides to growth accounting and public finance questions.
We recognize, of course, that although the economist should choose a welltested theory, every theory has some issues and questions that it does not
address well. In the case of neoclassical growth theory, for example, it fails
spectacularly when used to address economic development issues. Differences
in stocks of reproducible capital stocks cannot account for international
differences in per capita incomes. This does not preclude its usefulness in
evaluating tax policies and in business cycle research.
Construct a Model Economy
With a particular theoretical framework in mind, the third step in conducting
a computational experiment is to construct a model economy. Here, key
issues are the amount of detail and the feasibility of computing the equilibrium
process. Of-ten, economic experimenters are constrained to deal with a much
3
Neoclassical growth theory also represents a good example of the importance of interaction
between factual studies and theory development. Solow (1970) lists several growth facts that
influenced the development of neoclassical growth theory. Once the main ingredients of the
theory were established—such as the production function—new light was thrown on the data.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE COMPUTATIONAL EXPERIMENT
241
The Computational Experiment: An Econometric Tool 73
simpler model economy than they would like because computing the
equilibrium of a more complicated model would be impossible, given currently
available tools.
This situation is no different from that in the physical sciences, where, as in
economics, the computational experiment has become accepted as an
invaluable scientific tool. For example, in his overview of climate modeling,
Schneider (1987, p. 72) states: “Although all climate models consist of
mathematical representations of physical processes, the precise composition
of a model and its complexity depend on the problem it is designed to
address.” And later (p. 72): “Often it makes sense to attack a problem first
with a simple model and then employ the results to guide research at higher
resolution.” In the physical sciences, as in economics, confidence in a particular
framework or approach is gained through successful use.
So far, most of the model environments that macroeconomists have used
share certain characteristics. The environments are inhabited by a large number
of people whose decision problems are described explicitly. Both the household
and business sectors play a central role. For some questions, government or
foreign sectors must be included as well. In some models everyone is alike; in
others, such as those designed to address questions where demographic changes
are important, heterogeneous people must be used.
This description may sound somewhat indefinite or abstract, but we
reemphasize that an abstraction can be judged only relative to some given
question. To criticize or reject a model because it is an abstraction is foolish: all
models are necessarily abstractions. A model environment must be selected
based on the question being addressed. For example, heterogeneity of people is
crucial in the Auerbach and Kotlikoff (1987) model, which predicts the
consequences of the population’s changing age distribution on savings. However,
Ríos-Rull (1994) demonstrates that such life cycle features, even when combined
with elements of market incompleteness, are not quantitatively important to
business cycle findings regarding issues such as the contribution of technology
shocks to business cycle fluctuations. The features of a given model may be
appropriate for some question (or class of questions) but not for others.4
The selection and construction of a particular model economy should not
depend on the answer provided. In fact, searching within some parametric
class of economies for the one that best fits a set of aggregate time series
makes little sense, because it isn’t likely to answer an interesting question. For
example, if the question is of the type, “how much of fact X is accounted for
4
We will not debate the legitimacy of these methods, since such debates generally serve to
define schools rather than to produce agreement. Such debates are almost nonexistent during
normal science, but tend to recur during scientific revolutions. As stated by Kuhn (1962, p.
145), “Few philosophers of science still seek absolute criteria for the verification of scientific
theories.” All historically significant theories have agreed with the facts, but only to a degree.
No more precise answer can be found to the question of how well an individual theory fits the
facts. Using probabilistic verification theories that ask us to compare a given scientific theory
with all others that might fit the same data is a futile effort. We agree with Kuhn (p. 146) that
“probabilistic theories disguise the verification situation as much as they illuminate it.”
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
242
METHODOLOGY
74 Journal of Economic Perspectives
by Y,” then choosing the parameter values in such a way as to make the
amount accounted for as large as possible according to some metric is an
attempt to get a particular—not a good—answer to the question.
Calibrate the Model Economy
Now that a model has been constructed, the fourth step in carrying out a
computational experiment is to calibrate that model. Originally, in the physical
sciences, calibration referred to the graduation of measuring instruments. For
example, a Celsius thermometer is calibrated to register zero degrees when
immersed in water that contains ice and 100 degrees when immersed in boiling
water. A thermometer relies on the theory that mercury expands
(approximately) linearly within this range of temperatures. Related theory
also tells us how to recalibrate the thermometer if the measurements are
made in Denver or Mexico City rather than at sea level. In a sense, model
economies, like thermometers, are measuring devices. Generally, some economic
questions have known answers, and the model should give an approximately
correct answer to them if we are to have any confidence in the answer given
to the question with unknown answer. Thus, data are used to calibrate the
model economy so that it mimics the world as closely as possible along a
limited, but clearly specified, number of dimensions.
Note that calibration is not an attempt at assessing the size of something: it
is not estimation. Estimation is the determination of the approximate quantity
of something. To estimate a parameter, for example, a researcher looks for a
situation in which the signal-to-noise ratio is high. Using the existing data and
some theory, the researcher then constructs a probability model. An
estimator is developed that is robust, relative to the parameter that is to be
estimated, to the questionable features of the maintained hypothesis. As a
second example of estimation, a computational experiment itself is a type of
estimation, in the sense that the quantitative answer to a posed question is an
estimate. For example, the quantitative theory of a computational experiment
can be used to measure the welfare implications of alternative tax policies.
It is important to emphasize that the parameter values selected are not the
ones that provide the best fit in some statistical sense. In some cases, the
presence of a particular discrepancy between the data and the model
economy is a test of the theory being used. In these cases, absence of that
discrepancy is grounds to reject the use of the theory to address the question.
One such example is the use of standard theory to answer the question of
how volatile the postwar U.S. economy would have been if technology shocks
had been the only contributor to business cycle fluctuations. To answer this
question, a model economy with only technology shocks was needed. Using
the standard neoclassical production function, standard preferences to
describe people’s willingness to substitute intra- and intertemporally between
consumption and leisure, and an estimate of the technology shock variance,
we found that the model economy displays business cycle fluctuations 70
percent as large as did the U.S. economy (Kydland and Prescott, 1991b). This
number is our answer to the posed question.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE COMPUTATIONAL EXPERIMENT
Finn E.Kydland and Edward C.Prescott
243
75
Some have questioned our finding, pointing out that on one key
dimension real business cycle models and the world differ dramatically: the
correlation between hours worked and average labor productivity is near
one in the model economy and approximately zero in U.S. postwar
observations (McCallum, 1989). The detractors of the use of standard
theory to study business cycles are correct in arguing that the magnitude of
this correlation in the world provides a test of the theory. They are incorrect
in arguing that passing this test requires the value of this correlation in the
model and in the real world to be approximately equal. An implication of
the theory is that this correlation is a function of the importance of
technology shocks relative to other shocks. In particular, the less is the
relative importance of technology shocks, the smaller this correlation should
be. The reason for this is that the factors other than technology shocks that
give rise to variation in the labor input result in productivity being low when
hours are high.5 Given that the estimated contribution of technology shocks
to fluctuations is 70 percent, the correlation between hours and labor
productivity being near one in the data would have been grounds for
dismissing our answer. (For further elaboration on this point, see Kydland
and Prescott, 1991b, p. 79; Aiyagari, 1994.)
Run the Experiment
The fifth and final step in conducting a computational experiment is to run the
experiment. Quantitative economic theory uses theory and measurement to
estimate how big something is. The instrument is a computer program that
determines the equilibrium process of the model economy and uses this
equilibrium process to generate equilibrium realizations of the model economy.
The computational experiment, then, is the act of using this instrument. These
equilibrium realizations of the model economy can be compared with the
behavior of the actual economy in some period as follows.
If the model economy has no aggregate uncertainty, then it is simply a
matter of comparing the equilibrium path of the model economy with the
path of the actual economy. If the model economy has aggregate uncertainty,
first a set of statistics that summarize relevant aspects of the behavior of the
actual economy is selected. Then the computational experiment is used to
generate many independent realizations of the equilibrium process for the
model economy. In this way, the sampling distribution of this set of statistics
can be determined to any degree of accuracy for the model economy and
compared with the values of the set of statistics for the actual economy. In
comparing the sampling distribution of a statistic for the model economy to
the value of that statistic for the actual data, it is crucial that the same statistic
be computed for the model and the real world. If, for example, the statistic for
5
Christiano and Eichenbaum (1992a) have established formally this possibility in the case
where the other shock is variations in public consumption, but the result holds for any shock
that is approximately orthogonal to the technology shocks.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
244
METHODOLOGY
76 Journal of Economic Perspectives
the real world is for a 50-year period, then the statistic for the model economy
must also be for a 50-year period.
The Computational Experiment in Business Cycle Research
Business cycles, that is, the recurrent fluctuations of output and employment
about trend, puzzled economists for a long time. Understanding business cycles
required the development of methods that made possible the use of
computational experiments to answer questions concerned with the behavior
of dynamic economies with uncertainty. Prior to the development of these
methods, business cycle fluctuations were viewed as deviations from theory,
and very little progress was made in understanding them. Subsequent to the
development of those methods, computational experiments have been
extensively used in business cycle research. The results of these experiments
forced economists to revise their old views, and business cycle fluctuations are
now seen as being, in fact, predicted by standard theory. For these reasons, we
choose business cycle theory to illustrate the use of computational experiments
in economic research.
Posing Questions about the Business Cycle
In the 1970s, a common assumption behind research on the business cycle
was that one set of factors, most likely monetary shocks, was behind the
cyclical component and that an entirely different set of factors, mainly the
growth of productivity and inputs summarized by the neoclassical growth
model, accounted for the movement of the long-run growth component.
But there was also an earlier view, tracing as far back as work by Wicksell
(1907), that suggested that fluctuation in technological growth could produce
broad economic fluctuations. In the 1960s and 1970s, this source was
dismissed by many as being unlikely to play much of a role in the aggregate.
Most researchers believed that considerable variation could exist in
productivity at the industry level, but they believed that industry-level shocks
would average out in the aggregate. During the 1980s, however, technology
shocks gained renewed interest as a major source of fluctuations, supported
largely by computational experiments and quantitative economic theory. As a
consequence, business cycle theory treats growth and cycles as being
integrated, not as a sum of two components driven by different factors.6
6
An operational way of defining the trend empirically is described in Hodrick and Prescott
(1980), who used standard curve-fitting techniques to define a growth component as being
the curve that best fits a time series in a least-square sense, subject to a penalty on the sum
of the second differences squared. The larger this penalty parameter, the smoother the
fitted curve. For quarterly series, they found that a penalty parameter of 1600 made the
fitted curve mimic well the one that business cycle analysts would draw. Given the
finding that business cycle fluctuations are quantitatively just what neoclassical growth
theory predicts, the resulting deviations from trend are nothing more than well-defined
statistics. We emphasize that given the way the theory has developed, these statistics measure
nothing. Business cycle theory treats growth and cycles as being integrated, not as a sum of
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE COMPUTATIONAL EXPERIMENT
245
The Computational Experiment: An Econometric Tool 77
Thus, the fundamental question here is the extent to which neoclassical
growth theory can account for business cycle fluctuations, as well as long-term
growth trends. A particular question addressed was, “How much would the
U.S. postwar economy have fluctuated if technology shocks had been the
only source of fluctuations?” Computational experiments are well suited to
tackle this question.
The Theory Used in Model Selection
The basic theory used in the modern study of business cycles is the neoclassical
growth model. The basic version of this model can best be understood as
based on five relationships.
The first relationship is an aggregate production function that sets total
output equal to AtF(Kt, Ht), where F is a constant returns to scale function
where the inputs are capital and labor, and At is the technology level that
grows at random rates. In the simplest case, aggregate output is divided
between consumption C and investment I. Under the assumption that factors
are paid their marginal product, we obtain the identity that GNP and income
are equal: C+I=wH+rK, where w and r are factor rental prices.
The second relationship in the model economy describes the evolution of
the capital stock. Each time period, the existing capital stock depreciates at a
constant rate δ, but is replenished by new investment It. Thus Kt+1=(1–δ)Kt+It.
The third relationship describes the evolution of the technology parameter
At. Given that a structure that displays persistence is needed, a typical form
would be At+1=ρAt+⑀t+1, where ρ is large but less than one, and the shocks ⑀t+1
are identically and independently distributed. In other words, the technology
level for any given period depends on the technology level in the previous
period, plus a random disturbance. The technology described by these
relations specifies people’s ability to substitute.
The fourth relationship needed for a fully specified economy is a
specification of people’s willingness to substitute between consumption and
leisure, both intertemporally and intratemporally. For this purpose, our
model economy has a stand-in household with utility function that depends
on consumption and leisure. 7 For simplicity, one can assume that the
households own the capital stock directly and rent it to the firms.
two components driven by different factors. For that reason, talking about the resulting
statistics as imposing spurious cycles makes no sense. The Hodrick-Prescott filter is simply a
statistical decomposition that summarizes in a reasonable way what happens at business cycle
frequencies. The representation has proven useful in presenting the findings and judging the
reliability of the answer, as well as a way of demonstrating remaining puzzles or anomalies
relative to theory.
7
More explicitly, the function can take the general form
where we normalize so that market and nonmarket productive time add to one. For a
complete specification of the model, values of the parameters β, δ and ρ are needed, as well
as the explicit utility function U and the production function F.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
246
METHODOLOGY
78 Journal of Economic Perspectives
The final required element is an equilibrium concept. The one used is the
competitive equilibrium, which equates marginal rates of substitution and
transformation to price ratios. This involves equilibrium decision functions
for consumption, investment and labor input as functions of the capital stock
and productivity level during that time period: C(Kt, At), I(Kt, At) and H(Kt, At),
respectively. In other words, using dynamic programming methods, the
decisions can be computed as functions of the list of state variables that
provide sufficient information about the position of the economy.
Through this theory, business cycle theorists make contact with other fields
of economics. Macroeconomics is no longer largely separate from the rest of
economics. The utility and production functions used by business cycle theorists
are similar to those used by public finance researchers (for example, Auerbach
and Kotlikoff, 1987). The introduction of household production illustrates the
close connection with the work of labor economists (for example, Benhabib,
Rogerson and Wright, 1991; Greenwood and Hercowitz, 1991). When these
models are expanded to consider international trade explicitly, they draw upon
work in that field (Backus, Kehoe and Kydland, 1994).
The choice of a model, as already noted, must be governed both by the
question at hand and by what is computable.
As an example of altering the model to suit the posed question, consider a
contrast between two otherwise similar models. Benhabib, Rogerson and
Wright (1991) and Greenwood and Hercowitz (1991) both consider household
production in addition to market production, but the two studies are motivated
by somewhat different questions. Both use capital and labor as inputs in
nonmarket production. Benhabib and his coauthors divide the time allocation
of households into three uses: market and nonmarket production time and
leisure time. The model is designed to capture the household decision to
combine its labor with machines, such as stoves and washing machines, to
produce household consumption services. The authors argue that houses do
not need to be combined with labor, at least not to the same extent that
household machines do, so they exclude housing capital from their concept of
household capital. Greenwood and Hercowitz, on the other hand, distinguish
only between market and nonmarket time and include the stock of housing,
along with consumer durables, in their concept of household capital.8 This
example illustrates the important point that even the definition of particular
variables in relation to the model economy may depend on the question.
If a model environment is not computable, then it cannot be used for a
computational experiment. This restriction can be a severe one, and the
development of appropriate computable methods must therefore be given
high priority. Fortunately, there has been considerable progress in this area
over the last 30 or 40 years. In cases involving uncertain intertemporal
behavior, the development of statistical decision theory has provided a
8
To be consistent, they then subtract gross housing product (the measure of the service flow
from the economy’s housing stock) from GNP and add it to the consumer durables component of personal consumption expenditures.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE COMPUTATIONAL EXPERIMENT
Finn E.Kydland and Edward C.Prescott
247
79
consistent way for people to make decisions under uncertainty. Another
significant development is the Arrow-Debreu general equilibrium theory,
which extends equilibrium theory to uncertain environments.9 More recently,
Ríos-Rull (1995) offers an overview of the expansion in computable general
equilibrium models that incorporate heterogeneity in the household sector—a
category that has expanded dramatically over the last few years.
Calibration
Often, calibration involves the simple task of computing a few averages. For
example, if the standard Cobb-Douglas production function is used—that is, if
we let F(K, H) = K1-θHθ—then a numerical value for the parameter θ can be
obtained by computing the average labor share of total output over a period
of years. Several other growth relations map more or less directly into
parameter values for typical models within the neoclassical growth framework,
at least if the functional forms have been chosen with calibration in mind.
Most growth relations have not changed much, on average, from one cycle to
the next for several decades. As a consequence, computational experiments
replicate the key long-term or growth relations among model aggregates.
Exceptions do exist, where long-term relationships are not stable. For
example, the inventory stock as a fraction of G NP has declined steadily.
Durables expenditures as a fraction of total output have risen. For some
purposes these changes can be ignored, since that feature does not
significantly affect the answer to the question posed. At other times,
depending on the associated pattern in the corresponding relative price, such
information enables the researcher to obtain a quite precise estimate of some
elasticity of substitution, which can then be built into the model.
A good example of this sort of issue is the fact that hours of work per
household are about the same now as four decades ago, in spite of a large rise
in the real wage rate over the same period. This fact indicates that the
elasticity of substitution between consumption and nonmarket time is near
one. Still, many business cycle models abstract from the long-run productivity
growth that is required to imply this sort of wage growth, because the answer
to the questions addressed in those studies would have been essentially the
same, as shown by Hansen (1986).10
To calibrate a utility function for the household sector of the economy, it is
common to rely on averages across large numbers of the relevant population
in the actual economy. For example, some model environments employ a
utility function in consumption and leisure that, like the Cobb-Douglas
9
Also important is the development of recursive methods for the study of economic dynamics, because these methods allow economists to use the computational experiment to generate
time series disciplined by factual studies (Stokey and Lucas, 1989).
10
Hansen (1986) compares otherwise identical model economies and permits growth in one
version and not in the other. The model without growth needs a slight adjustment in the
capital depreciation rate to be calibrated to the investment share of output and the observed
capital/output ratio. With this adjustment, both models estimate the same role of technological shocks (more precisely, Solow residuals) for cyclical fluctuations.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
248
80
METHODOLOGY
Journal of Economic Perspectives
production function above, has a share parameter. In this case, the weight that
should be placed on consumption turns out to be approximately equal to the
average fraction of time spent in market activity. This fraction, in principle,
can be obtained from panel data for large samples of individuals. Ghez and
Becker (1975) offer a careful measurement study—making reasonable and
thoughtful judgments about factors like age limits of the population sample
and definition of total time allocated to market and nonmarket activities,
including treatment of sleep and personal care.
In calibration, we sometimes make the model economy inconsistent with
the data on one dimension so that it will be consistent on another. For
example, Imrohoroglu (1992) explores the welfare consequences of
alternative monetary arrangements in worlds where agents are liquidity
constrained, while Cooley and Hansen (1989) explore the welfare
consequences in worlds where people use money for transaction purposes.
These are two very different environments, each of which abstracts from the
main feature of the other. Imrohoroglu calibrates her model economy to yield
a stock of money held per household that is in line with U.S. observations. In
her model, however, people hold money because they do not have access to
an insurance technology to insure against randomness in the market value of
their time. Equivalently, if they do have access to such an insurance
technology, they find it so costly that, in equilibrium, they do not employ it.
This is the only reason, in her model, for people to hold money; if she had
calibrated the model to the amount of variation in individual income found in
panel data, the model would have implied that average household holdings of
liquid assets were about half of those actually held.
Of course, households have other reasons for holding liquid assets that earn
much less than the average return on physical capital. For instance, such assets
can be used as a down payment on a house at some future date, as a substitute
for insurance against sickness and accidents, or for transaction purposes, as in
the Cooley and Hansen (1989) environment. These and other factors are
abstracted from in the world of Imrohoroglu (1992), which led her to introduce
greater variation in the market value of households’ time so as to make per
capita holdings of money in the model match actual holdings. This calibration is
reasonable, given the question she addresses. Her implicit assumption is that it
is unimportant which liquidity factor gives rise to these holdings. Subsequent
research will either support this working hypothesis or disprove it and, in the
process, lead to better model economies for evaluating monetary and credit
policy arrangements. This sequence is how economic science progresses.
Running Experiments
With explicit functional forms for the production and utility functions, with
values assigned to the parameters, and with a probability distribution for the
shocks, a researcher can use the model economy to perform computational
experiments. The first step is to compute the equilibrium decision rules, which
are functions of the state of the economy. The next step is to generate
equilibrium realizations of the economy. The computer begins each period
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE COMPUTATIONAL EXPERIMENT
249
The Computational Experiment: An Econometric Tool 81
with a given level of the state variables, for example, the capital stock and the
technology level. The values of the state variables along with the equilibrium
decision and pricing functions determine the equilibrium realization for that
period. Equilibrium investment and the new technology shocks determine
next period’s state. In the next and subsequent periods, this procedure is
repeated until time series of the desired length are obtained. The resulting
model time series can then be summarized by a suitable set of statistics.
In Kydland and Prescott (1982), we built a model economy where all
fluctuations could be traced back to technology shocks. We began by extending
the neoclassical growth model to include leisure as an argument of the stand-in
household’s utility function. Given that more than half of business cycle
fluctuations are accounted for by variations in the labor input, introducing this
element was crucial. We then calibrated the deterministic version of the model
so that its consumption-investment shares, factor income shares, capital/output
ratios, leisure/market-time shares and depreciation shares matched the average
values for the U.S. economy in the postwar period. We abstracted from public
finance considerations and consolidated the public and private sectors. We
introduced Frisch’s (1933b) assumption of time to build new productive capital.
The construction period considered was four quarters, with new capital
becoming productive only upon completion, but with resources being used up
throughout the construction period. Given the high volatility of inventory
investment, inventory stocks were included as a factor of production. In our
computational experiments, using technology shock variance estimated from
production function residuals (Prescott, 1986), we found that the model
economy’s output variance was 55 percent as large as the corresponding
variance for the U.S. economy in the postwar period.
Probably the most questionable assumption of this theory is that of
homogeneous workers, with the additional implication that all variation in hours
occurs in the form of changes in hours per worker. According to aggregate data
for the U.S. economy, only about one-third of the quarterly fluctuations in
market hours are of this form, while the remaining two-thirds arise from changes
in the number of workers (Kydland and Prescott, 1990, Table 1).
This observation led Hansen (1985) to introduce the Rogerson (1988)
labor indivisibility construct into a business cycle model. In the Hansen world,
all fluctuations in hours are in the form of employment variation. To deal with
the apparent nonconvexity arising from the assumption of indivisible labor,
Hansen makes the problem convex by assuming that the commodity points
are contracts in which every agent is paid the same amount whether that
agent works or not, and that a lottery randomly chooses who actually works
in every period. He finds that with this labor indivisibility, his model economy
fluctuates as much as did the U.S. economy. Our view is that with the
extreme assumption of fluctuations only in employment, Hansen
overestimates the amount of aggregate fluctuations accounted for by Solow
residuals in the same way that our equally extreme assumption of fluctuations
solely in hours per worker led to an underestimation.
In Kydland and Prescott (1991b), the major improvement on the 1982
version of the model economy is that variation is permitted in both the
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
250
82
METHODOLOGY
Journal of Economic Perspectives
number of workers and the number of hours per worker. The number of
hours in which a plant is operated in any given period is endogenous.
Because the cost of moving workers in and out of the labor force is not
included, a property of the equilibrium is that all of the hours variation is in
the form of employment change and none in hours per worker. In that
respect, the Kydland and Prescott (1991b) model is identical to Hansen’s
(1985) model. Using the economy with no moving costs, technology shocks
are estimated to account for about 90 percent of the aggregate output
variance. For the economy with moving costs, we calibrated it so that relative
variations in hours per worker and the number of workers matched U.S.
data. The estimate of the fraction of the cycle accounted for by technology
shocks is then reduced to 70 percent.
These estimates of the contribution of technology shocks to aggregate
fluctuations have been found to be robust in several modifications of the model
economy. For instance, Greenwood, Hercowitz and Huffman (1988) permit the
utilization rate of capital to vary and affect its depreciation rate and assume all
technology change is embodied in new capital. Danthine and Donaldson (1990)
introduce an efficiency-wage construct, while Cho and Cooley (1995) permit
nominal-wage contracting. Ríos-Rull (1994) uses a model calibrated to life cycle
earnings and consumption patterns. Gomme and Greenwood (1995)
incorporate heterogeneous agents with recursive preferences and equilibrium
risk allocations. In none of these cases is the estimate of the contribution of
technology shocks to aggregate fluctuations significantly altered.
The computational experiment is also being used to derive the quantitative
implications of monetary shocks for business cycle fluctuations if money is
used for transaction purposes only. In these experiments, money may be held
either because it is required in advance of purchasing cash goods (Lucas and
Stokey, 1987; Cooley and Hansen 1989, 1992) or because real cash balances
economize on time (Kydland, 1989). Models of this type have been used to
evaluate monetary policy.
At this stage, we are less confident in these model economies than those
designed to evaluate the contribution of technology shocks. There are three
related reasons. The first is that, unlike actual economies, these model
economies fail to display the sluggishness of the inflation rate’s response to
changes in the growth rate of money (Christiano and Eichenbaum, 1992b). The
second is that people seem to hold far larger monetary assets than are needed
for transaction purposes—in the postwar period, for example, U.S. households’
holdings of M2 have exceeded half of annual GNP—which implies that the
transaction rationale for holding money is not well understood. The third
reason is that the evaluation of monetary policy appears to be sensitive to the
reason people hold these liquid assets. Imrohoroglu (1992) has constructed a
model economy in which people vary their holdings of liquid assets as their
income varies to smooth their consumption.11 She finds that if a transaction-cost
model is calibrated to data generated by her model economy and the calibrated
11
Imrohoroglu and Prescott (1991) introduce a banking technology to intermediate government debt.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE COMPUTATIONAL EXPERIMENT
Finn E.Kydland and Edward C.Prescott
251
83
economy is used to estimate the cost of inflation, this estimate would grossly
underestimate the true cost of inflation for her model world. This result is
surprising and bothersome. Typically, how some feature is introduced is
unimportant as long as the aggregate substitution elasticities and quantities
match. We currently do not have the tools for computing equilibria of models
with both the features of the neoclassical growth model and the idiosyncratic
shocks that result in people holding money for precautionary reasons. One
may say that stronger theory is needed when it comes to evaluating the
contribution of monetary policy shocks to business cycle fluctuations.
Summary
With the general equilibrium approach, empirical knowledge is organized around
preferences and technologies. Given the question and given existing economic
theory and measurement, a researcher creates a model economy. This
researcher then determines a quantitative answer to the posed question for
the model economy. If the theory is strong and the measurements good, we
have confidence that the answer for the model economy will be essentially the
same as for the actual economy.
Of course, sometimes measurement is not very good, and a series of
computational experiments reveals that different plausible values of some
parameter give very different answers to the posed question. If so, this
parameter—which measures some aspect of people’s willingness and ability to
substitute—must be more accurately measured before theory can provide an
answer in which we have confidence. Or sometimes the theory relative to the
question is weak or nonexistent, and the answer depends upon which of the
currently competing theories is used to construct the model economy. If so,
these competing theories must be subjected to further tests before there is a
good basis for choosing among them. At still other times, the computational
tools needed to derive the implications of the theory do not exist, so better
computational methods or more powerful computers are needed.
Earlier in this article, we quoted the Lucas (1980) definition of “theory” as
being an explicit set of instructions for building an imitation economy to
address certain questions and not a collection of assertions about the behavior
of the actual economy. Consequently, statistical hypothesis testing, which is
designed to test assertions about actual systems, is not an appropriate tool for
testing economic theory.
One way to test a theory is to determine whether model economies
constructed according to the instructions of that theory mimic certain aspects
of reality. Perhaps the ultimate test of a theory is whether its predictions are
confirmed—that is, did the actual economy behave as predicted by the model
economy, given the policy rule selected? If a theory passes these tests, then it
is tested further, often by conducting a computational experiment that
includes some feature of reality not previously included in the computational
experiments. More often than not, introducing this feature does not change
the answers, and currently established theory becomes stronger. Occasionally,
however, the new feature turns out to be important, and established theory
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
252
84
METHODOLOGY
Journal of Economic Perspectives
must then be expanded and improved. In this way, economic science
progresses.
Given the infeasibility of controlled experiments with national economies,
the computational experiment is the tool of quantitative economic theory,
whether the primary concern be with theory use or with theory development
and testing.
ƒ We thank Graham Candler, Javier Díaz-Giménez, Tryphon Kollintzas, Jim Schmitz and Nancy Stokey
for helpful discussions and the National Science Foundation for financial support. The views expressed
herein are those of the authors and not necessarily those of the Federal Reserve Banks of Cleveland or
Minneapolis, nor of the Federal Reserve System.
References
Aiyagari, S.Rao, “On the Contribution of
Technology Shocks to Business Cycles,”
Quarterly Review, Federal Reserve Bank of
Minneapolis, Winter 1994, 18, 22–34.
Auerbach, Alan J., and Laurence
J.Kotlikoff, Dynamic Fiscal Policy. Cambridge,
U.K.: Cambridge University Press, 1987.
Backus, David K., Patrick J.Kehoe, and
Finn E.Kydland , “Dynamics of the Trade
Balance and the Terms of Trade: The J-curve?,”
American Economic Review, March 1994, 84,
84–103.
Benhabib, Jess, Richard Rogerson, and
Randall Wright, “Homework in
Macroeconomics: Household Production
and Aggregate Fluctuations,” Journal of Political
Economy, December 1991, 99, 1166–87.
Brown, Drussilla K., Alan V.Deardorff,
and Robert M.Stern, “Estimates of a North
American Free Trade Agreement,”
manuscript, Federal Reserve Bank of
Minneapolis, 1994.
Chang, Ly-June, “Business Cycles with
Distorting Taxes and Disaggregate Capital
Markets,” Journal of Economic Dynamics and
Control, July/September 1995, 19, 985–1010.
Cho, Jong Ok, and Thomas F.Cooley,
“The Business Cycle with Nominal Contracts,”
Economic Theory, June 1995, 6, 13–34.
Christiano, Lawrence J., and Martin
Eichenbaum, “Current Real-Business-Cycle
Theories and Aggregate Labor-Market
Fluctuations,” American Economic Review, June
1992a, 82, 430–50.
Christiano, Lawrence J., and Martin
Eichenbaum, “Liquidity Effects, Monetary
Policy, and the Business Cycle.” Working
Paper No. 4129, National Bureau of
Economic Research, 1992b.
Cooley, Thomas F., and Gary
D.Hansen, “The Inflation Tax in a Real
Business Cycle Model,” American Economic
Review, September 1989, 79, 733–48.
Cooley, Thomas F., and Gary
D.Hansen, “Tax Distortions in a Neoclassical
Monetary Economy,” Journal of Economic
Theory, December 1992, 58, 290–316.
Danthine, Jean-Pierre, and John
B.Donaldson, “Efficiency Wages and the
Business Cycle Puzzle,” European Economic
Review, November 1990, 34, 1275–1301.
Finn, Mary G., “Variance Properties of
Solow’s Productivity Residual and their
Cyclical Implications,” Journal of Economic
Dynamics and Control, July/September 1995,
19, 1249–82.
Frisch, Ragnar, “Editorial,” Econometrica,
January 1933a, 1, 1–5.
Frisch, Ragnar, “Propagation Problems
and Impulse Problems in Dynamic
Economics.” In Economic Essays in Honour of
Gustav Cassel. London: G.Allen & Unwin Ltd.,
1933b, pp. 171–205.
Ghez, Gilbert R., and Gary S.Becker, The
Allocation of Time and Goods over the Life Cycle.
New York: Columbia University Press, 1975.
Gomme, Paul, and Jeremy Greenwood,
“On the Cyclical Allocation of Risk,” Journal
of Economic Dynamics and Control, January/
February 1995, 19, 91–124.
Gravelle, Jane G., and Laurence
J.Kotlikoff, “Corporate Taxation and the
Efficiency Gains of the 1986 Tax Reform
Act,” Economic Theory, June 1995, 6, 51–82.
Greenwood, Jeremy, and Zvi Hercowitz,
“The Allocation of Capital and Time Over
the Business Cycle,” Journal of Political Economy,
December 1991, 99, 1188–214.
Greenwood, Jeremy, Zvi Hercowitz, and
Gregory W.Huffman, “Investment,
Capacity Utilization, and the Real Business
Cycle,” American Economic Review, June 1988,
78, 402–18.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
THE COMPUTATIONAL EXPERIMENT
253
The Computational Experiment: An Econometric Tool 85
Gregory, Allan W., and Gregory
W.Smith, “Statistical Aspects of Calibration
in Macroeconomics.” In Maddala, G.S.,
C.R.Rao, and H.D.Vinod, eds., Handbook of
Statistics. Vol. 2, New York: Elsevier Science,
1993, pp. 703–19.
Hansen, Gary D., “Indivisible Labor and
the Business Cycle,” Journal of Monetary
Economics, November 1985, 16, 309–27.
Hansen, Gary D., “Growth and
Fluctuations.” In “Three Essays on Labor
Indivisibility and the Business Cycle,” Ph.D.
dissertation, University of Minnesota, 1986.
Harris, Richard, “Applied General
Equilibrium Analysis of Small Open
Economies with Scale Economies and
Imperfect Competition,” American Economic
Review, December 1984, 18, 1016–32.
Hodrick, Robert J., and Edward
C.Prescott, “Post-War U.S. Business Cycles:
An Empirical Investigation.” Discussion Paper
No. 451, Carnegie-Mellon University, 1980.
Hopenhayn, Hugo, and Richard
Rogerson, “Job Turnover and Policy
Evaluation: A General Equilibrium Analysis,”
Journal of Political Economy, October 1993, 101,
915–38.
Hornstein, Andreas, “Monopolistic
Competition, Increasing Returns to Scale and
the Importance of Productivity Change,”
Journal of Monetary Economics, June 1993, 31,
299–316.
Imrohoroglu, Ayse, “The Welfare Cost of
Inflation Under Imperfect Insurance,” Journal
of Economic Dynamics and Control, January 1992,
16, 79–91.
Imrohoroglu, Ayse, and Edward
C.Prescott, “Seigniorage as a Tax: A
Quantitative Evaluation, “Journal of Money,
Credit and Banking, August 1991, 23, 462–75.
Kuhn, Thomas S., The Structure of
Scientific Revolutions. Chicago: University of
Chicago Press, 1962.
Kydland, Finn E., “The Role of Money in
a Business Cycle Model.” Discussion Paper No.
23, Institute for Empirical Macroeconomics,
Federal Reserve Bank of Minneapolis, 1989.
Kydland, Finn E., and Edward
C.Prescott, “Time to Build and Aggregate
Fluctuations,” Econometrica, November 1982,
50, 1345–70.
Kydland, Finn E., and Edward C.Prescott,
“Business Cycles: Real Facts and a Monetary
Myth,” Quarterly Review, Federal Reserve Bank of
Minneapolis, Spring 1990, 14, 3–18.
Kydland, Finn E., and Edward
C.Prescott, “The Econometrics of the
General Equilibrium Approach to Business
Cycles,” Scandinavian Journal of Economics,
1991a, 93:2, 161–78.
Kydland, Finn E., and Edward
C.Prescott, “Hours and Employment
Variation in Business Cycle Theory,” Economic
Theory, January 1991b, 1, 63–81.
Lucas, Robert E., Jr., “Methods and
Problems in Business Cycle Theory,” Journal
of Money, Credit and Banking, November
1980, 12, 696–715.
Lucas, Robert E., Jr., and Nancy
L.Stokey, “Money and Interest in a Cash-inAdvance Economy,” Econometrica, May 1987,
55, 491–513.
Marschak, Jacob, “Introduction to
Econometrics,” hectographed lecture notes,
University of Buffalo, 1948.
McCallum, Bennett T., “Real Business
Cycle Models.” In Barro, Robert J., ed., Modern
Business Cycle Theory. Cambridge, Mass.:
Harvard University Press, 1989, pp. 16–50.
Prescott, Edward C., “Theory Ahead of
Business Cycle Measurement,” Quarterly Review,
Federal Reserve Bank of Minneapolis, Fall 1986,
10, 9–22; also in Brunner, Karl, and Allan
H.Meltzer, eds., Real Business Cycles, Real Exchange
Rates, and Actual Policies, Carnegie-Rochester
Conference Series on Public Policy. Vol. 25,
Amsterdam: North-Holland, 1986, pp. 11–44.
Ríos-Rull, José-Víctor, “On the
Quantitative Importance of Market
Completeness,” Journal of Monetary Economics,
December 1994, 34, 462–96.
Ríos-Rull, José-Víctor, “Models with
Heterogeneous Agents.” In Cooley, T.F., ed.,
Frontiers of Business Cycle Research. Princeton,
N.J.: Princeton University Press, 1995.
Rogerson, Richard, “Indivisible Labor,
Lotteries and Equilibrium,” Journal of
Monetary Economics, January 1988, 21, 3–16.
Schneider, Stephen H., “Climate
Modeling,” Scientific American, May 1987, 256,
72–80.
Shoven, John B., and J.Whalley, “A
General Equilibrium Calculation of the
Differential Taxation of Income from Capital
in the U.S.,” Journal of Public Economics,
November 1972, 1, 281–321.
Solow, Robert M., Growth Theory: An
Exposition. Oxford: Clarendon Press, 1970.
Stokey, Nancy L., and Robert E.Lucas,
Jr. (with Edward C.Prescott), Recursive Methods
in Economic Dynamics. Cambridge, Mass.:
Harvard University Press, 1989.
Wicksell, Knut, “Krisernas gåta,”
Statsøkonomisk Tidsskrift, 1907, 21, 255–70.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
254
CHAPTER 14
Journal of Economic Perspectives—Volume 10, Number 1—Winter 1996—Pages 87–104
The Empirical Foundations of
Calibration
Lars Peter Hansen and James J.Heckman
G
eneral equilibrium theory provides the intellectual underpinnings for
modern macroeconomics, finance, urban economics, public finance
and numerous other fields. However, as a paradigm for organizing and
synthesizing economic data, it poses some arduous challenges. A widely
accepted empirical counterpart to general equilibrium theory remains to be
developed. There are few testable implications of general equilibrium theory
for a time series of aggregate quantities and prices. There are a variety of ways
to formalize this claim. Sonnenschein (1973) and Mantel (1974) show that
excess aggregate demand functions can have “arbitrary shapes” as long as
Walras’ Law is satisfied. Similarly, Harrison and Kreps (1979) show that a
competitive equilibria can always be constructed to rationalize any arbitrage-free
specification of prices. Observational equivalence results are pervasive in
economics. There are two responses to this state of affairs. One can view the
flexibility of the general equilibrium paradigm as its virtue. Since it is hard to
reject, it provides a rich apparatus for interpreting and processing data. 1
Alternatively, general equilibrium theory can be dismissed as being empirically
irrelevant because it imposes no testable restrictions on market data.
Even if we view the “flexibility” of the general equilibrium paradigm as a
virtue, identification of preferences and technology is problematic. Concern
1
Lucas and Sargent (1988) make this point in arguing that early Keynesian critiques of
classical economics were misguided by their failure to recognize this flexibility.
ƒ Lars Peter Hansen is Homer J.Livingston Professor of Economics, and James Heckman is Henry
Schultz Distinguished Service Professor of Economics and Director of the Center for Social Program
Evaluation at the Irving B.Harris School of Public Policy Studies, all at the University of Chicago,
Chicago, Illinois.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
EMPIRICAL FOUNDATIONS OF CALIBRATION
88
255
Journal of Economic Perspectives
about the lack of identification of aggregate models has long troubled
econometricians (for example, Liu, 1960; Sims, 1980). The tenuousness of
identification of many models makes policy analysis and the evaluation of the
welfare costs of programs a difficult task and leads to distrust of aggregate
models. Different models that “fit the facts” may produce conflicting estimates
of welfare costs and dissimilar predictions about the response of the economy
to changes in resource constraints.
Numerous attempts have been made to circumvent this lack of identification,
either by imposing restrictions directly on aggregate preferences and technologies,
or by limiting the assumed degree of heterogeneity in preferences and
technologies. For instance, the constant elasticity of substitution specification for
preferences over consumption in different time periods is one of workhorses of
dynamic stochastic equilibrium theory. When asset markets are sufficiently rich, it
is known from Gorman (1953) that these preferences can be aggregated into the
preferences of a representative consumer (Rubinstein, 1974). Similarly, CobbDouglas aggregate production functions can be obtained from Leontief micro
technologies aggregated by a Pareto distribution for micro productivity
parameters (Houthakker, 1956). These results give examples of when simple
aggregate relations can be deduced from relations underlying the micro behavior
of the individual agents, but they do not justify using the constructed aggregate
relations to evaluate fully the welfare costs and benefits of a policy.2
Micro data offer one potential avenue for resolving the identification problem,
but there is no clear formal statement that demonstrates how access to such data
fully resolves it. At an abstract level, Brown and Matzkin (1995) show how to use
information on individual endowments to obtain testable implications in exchange
economies. As long as individual income from endowments can be decomposed
into its component sources, they show that the testability of general equilibrium
theory extends to production economies. Additional restrictions and considerable
price variation are needed to identify microeconomic preference relations for data
sets that pass the Brown-Matzkin test.
Current econometric practice in microeconomics is still far from the
nonparametric ideal envisioned by Brown and Matzkin (1995). As shown by
Gorman (1953), Wilson (1968), Aigner and Simon (1970) and Simon and
Aigner (1970), it is only under very special circumstances that a micro
parameter such as the intertemporal elasticity of substitution or even a
marginal propensity to consume out of income can be “plugged into” a
representative consumer model to produce an empirically concordant
aggregate model. As illustrated by Houthakker’s (1956) result,
microeconomic technologies can look quite different from their aggregate
counterparts. In practice, microeconomic elasticities are often estimated by
reverting to a partial equilibrium econometric model. Cross-market price
2
Gorman’s (1953) results provide a partial justification for using aggregate preferences to
compare alternative aggregate paths of the economy. Even if one aggregate consumptioninvestment profile is preferred to another via this aggregate preference ordering, to convert
this into a Pareto ranking for the original heterogeneous agent economy requires computing
individual allocations for the path—a daunting task.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
256
METHODOLOGY
Lars Peter Hansen and James J.Heckman
89
elasticities are either assumed to be zero or are collapsed into constant terms
or time dummies as a matter of convenience. General equilibrium,
multimarket price variation is typically ignored in most microeconomic
studies.
Battle lines are drawn over the issue of whether the microeconometric
simplifications commonly employed are quantitatively important in
evaluating social welfare and assessing policy reforms. Shoven and Whalley
(1972, 1992) attacked Harberger’s use of partial equilibrium analysis in
assessing the effects of taxes on outputs and welfare. Armed with Scarf’s
algorithm (Scarf and Hansen, 1973), they computed fundamentally larger
welfare losses from taxation using a general equilibrium framework than
Harberger computed using partial equilibrium analysis. However, these and
other applications of general equilibrium theory are often greeted with
skepticism by applied economists who claim that the computations rest on
weak empirical foundations. The results of many simulation experiments are
held to be fundamentally implausible because the empirical foundations of the
exercises are not secure.
Kydland and Prescott are to be praised for taking the general equilibrium
analysis of Shoven and Whalley one step further by using stochastic general
equilibrium as a framework for understanding macroeconomics.3 Their vision
is bold and imaginative, and their research program has produced many
creative analyses. In implementing the real business cycle program,
researchers deliberately choose to use simple stylized models both to
minimize the number of parameters to be “calibrated” and to facilitate
computations.4 This decision forces them to embrace a rather different notion
of “testability” than used by the other general equilibrium theorists, such as
Sonnenschein, Mantel, Brown and Matzkin. At the same time, the real
business cycle community dismisses conventional econometric testing of
parametric models as being irrelevant for their purposes. While Kydland and
Prescott advocate the use of “well-tested theories” in their essay, they never
move beyond this slogan, and they do not justify their claim of fulfilling this
criterion in their own research. “Well tested” must mean more than “familiar”
or “widely accepted” or “agreed on by convention,” if it is to mean anything.
Their suggestion that we “calibrate the model” is similarly vague. On one
hand, it is hard to fault their advocacy of tightly parameterized models,
because such models are convenient to analyze and easy to understand.
Aggregate growth coupled with uncertainty makes nonparametric
identification of preferences and technology extremely difficult, if not
impossible. Separability and homogeneity restrictions on preferences and
technologies have considerable appeal as identifying assumptions. On the
other hand, Kydland and Prescott never provide a coherent framework for
3
The earlier work by Lucas and Prescott (1971) took an initial step in this direction by providing
a dynamic stochastic equilibrium framework for evaluating empirical models of investment.
4
The term “real business cycle” originates from an emphasis on technology shocks as a
source of business cycle fluctuations. Thus, real, as opposed to nominal, variables drive the
process. In some of the recent work, both real and nominal shocks are used in the models.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
EMPIRICAL FOUNDATIONS OF CALIBRATION
90
257
Journal of Economic Perspectives
extracting parameters from microeconomic data. The same charge of having
a weak empirical foundation that plagued the application of deterministic
general equilibrium theory can be lodged against the real business cycle
research program. Such models are often elegant, and the discussions
produced from using them are frequently stimulating and provocative, but
their empirical foundations are not secure. What credibility should we attach
to numbers produced from their “computational experiments,” and why
should we use their “calibrated models” as a basis for serious quantitative
policy evaluation? The essay by Kydland and Prescott begs these fundamental
questions.
The remainder of our essay is organized as follows. We begin by
discussing simulation as a method for understanding models. This method is
old, and the problems in using it recur in current applications. We then
argue that model calibration and verification can be fruitfully posed as
econometric estimation and testing problems. In particular, we delineate the
gains from using an explicit econometric framework. Following this
discussion, we investigate current calibration practice with an eye toward
suggesting improvements that will make the outputs of computational
experiments more credible. The deliberately limited use of available
information in such computational experiments runs the danger of making
many economic models with very different welfare implications compatible
with the evidence. We suggest that Kydland and Prescott’s account of the
availability and value of micro estimates for macro models is dramatically
overstated. There is no filing cabinet full of robust micro estimates ready to
use in calibrating dynamic stochastic general equilibrium models. We outline
an alternative paradigm that, while continuing to stress the synergy between
microeconometrics and macro simulation, will provide more credible inputs
into the computational experiments and more accurate assessments of the
quality of the outputs.
Simulation as a Method for Understanding Models
In a simple linear regression model, the effect of an independent variable on
the dependent variable is measured by its associated regression coefficient. In
the dynamic nonlinear models used in the Kydland-Prescott real business
cycle research program, this is no longer true. The dynamic nature of such
models means that the dependent variable is generated in part from its own
past values. Characterizing the dynamic mechanisms by which exogenous
impulses are transmitted into endogenous time series is important to
understanding how these models induce fluctuations in economic aggregates.
Although there is a reliance on linearization techniques in much of the current
literature, for large impulses or shocks, the nonlinear nature of such models is
potentially important. To capture the richness of a model, the analyst must
examine various complicated features of it. One way to do this is to simulate
the model at a variety of levels of the forcing processes, impulses and
parameters.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
258
METHODOLOGY
The Empirical Foundations of Calibration
91
The idea of simulating a complex model to understand its properties is not
a new principle in macroeconomics. Tinbergen’s (1939) simulation of his
League of Nations model and the influential simulations of Klein and
Goldberger (1955) and Goldberger (1959) are but three of a legion of
simulation exercises performed by previous generations of economists.5 Fair
(1994) and Taylor (1993) are recent examples of important studies that rely
on simulation to elucidate the properties of estimated models.
However, the quality of any simulation is only as good as the input on which
it is based. The controversial part of the real business cycle simulation program
is the method by which the input parameters are chosen. Pioneers of simulation
and of economic dynamics like Tinbergen (1939) and Frisch (1933) often
guessed at the parameters they used in their models, either because the data
needed to identify the parameters were not available, or because the
econometric methods were not yet developed to fully exploit the available data
(Morgan, 1990). At issue is whether the state of the art for picking the
parameters to be used for simulations has improved since their time.
Calibration versus Estimation
A novel feature of the real business cycle research program is its endorsement
of “calibration” as an alternative to “estimation.” However, the distinction
drawn between calibrating and estimating the parameters of a model is artificial
at best.6 Moreover, the justification for what is called “calibration” is vague
and confusing. In a profession that is already too segmented, the construction
of such artificial distinctions is counterproductive. It can only close off a
potentially valuable dialogue between real business cycle research and other
research in modern econometrics.
Since the Kydland-Prescott essay is vague about the operating principles of
calibration, we turn elsewhere for specificity. For instance, in a recent description
of the use of numerical models in the earth sciences, Oreskes, Shrader-Frechette
and Belitz (1994, pp. 642, 643) describe calibration as follows:
In earth sciences, the modeler is commonly faced with the inverse
problem: The distribution of the dependent variable (for example, the
hydraulic head) is the most well known aspect of the system; the
distribution of the independent variable is the least well known. The
process of tuning the model—that is, the manipulation of the independent
variables to obtain a match between the observed and simulated
distribution or distributions of a dependent variable or variables—is known
as calibration.
5
Simulation is also widely used in physical science. For example, it is customary in the studies
of fractal dynamics to simulate models in order to gain understanding of the properties of
models with various parameter configurations (Peitgen and Richter, 1986).
6
As best we can tell from their essay, Kydland and Prescott want to preserve the term
“estimation” to apply to the outputs of their computational experiments.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
EMPIRICAL FOUNDATIONS OF CALIBRATION
92
259
Journal of Economic Perspectives
Some hydrologists have suggested a two-step calibration scheme in which
the available dependent data set is divided into two parts. In the first step,
the independent parameters of the model are adjusted to reproduce the
first part of the data. Then in the second step the model is run and the
results are compared with the second part of the data. In this scheme, the
first step is labeled “calibration” and the second step is labeled “verification.”
This appears to be an accurate description of the general features of the
“calibration” method advocated by Kydland and Prescott. For them, data for
the first step come from micro observations and from secular growth observations
(see also Prescott, 1986a). Correlations over time and across variables are to
be used in the second step of verification.
Econometricians refer to the first stage as estimation and the second stage as
testing. As a consequence, the two-stage procedure described by Oreskes,
Shrader-Frechette and Belitz (1994) has a straightforward econometric
counterpart.7
From this perspective, the Kydland-Prescott objection to mainstream
econometrics is simply a complaint about the use of certain loss functions for
describing the fit of a model to the data or for producing parameter estimates.
Their objection does not rule out econometric estimation based on other loss
functions.
Econometric estimation metrics like least squares, weighted least squares
or more general method-of-moments metrics are traditional measures of fit.
Difference among these methods lie in how they weight various features of
the data; for example, one method might give a great deal of weight to distant
outliers or to certain variables, causing them to pull estimated trend lines in
their direction; another might give less weight to such outliers or variables.
Each method of estimation can be justified by describing the particular loss
function that summarizes the weights put on deviations of a model’s
predictions from the data. There is nothing sacred about the traditional loss
functions in econometrics associated with standard methods, like ordinary
least squares. Although traditional approaches do have rigorous justifications,
a variety of alternative loss functions could be explored that weight particular
features of a model more than others. For example, one could estimate with a
loss function that rewarded models that are more successful in predicting
turning points. Alternatively, particular time series frequencies could be
deemphasized in adopting an estimation criterion because misspecification of
a model is likely to contaminate some frequencies more than others
(Dunsmuir and Hannan, 1978; Hansen and Sargent, 1993; Sims, 1993).
7
See Christiano and Eichenbaum (1992) for one possible econometric implementation of
this two-step approach. They use a generalized method of moments formulation (for example, Hansen, 1982) in which parameters are estimated by a first stage, exactly identified set
of moment relations, and the model is tested by looking at another set of moment restrictions.
Not surprisingly, to achieve identification of the underlying set of parameters, they are
compelled to include more than just secular growth relations in the first-stage estimation,
apparently violating one of the canons of current calibration practice.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
260
METHODOLOGY
Lars Peter Hansen and James J.Heckman
93
The real business cycle practitioners adopt implicit loss functions. In
looking at economic aggregates, their implicit loss functions appear to focus on
the model predictions for long-run means, to the exclusion of other features
of the data, when selecting parameter estimates. It is unfortunate that we are
forced to guess about the rationale for the loss functions implicit in their
research. There is little emphasis on assessing the quality of the resulting
calibration. Formalizing the criteria for calibration and verification via loss
functions makes the principle by which a particular model is chosen easier to
understand. A clear statement would lead to more fruitful and focused
conversations about the sources and reliability of estimated parameters.
As Oreskes, Shrader-Frechette and Belitz (1994) emphasize, the distinction
between calibration and verification is often contrived. In many circumstances
the verification step should really be considered part of the “calibration” step.
The absence of a sharp distinction between these two stages is consistent with
the difficulty of obtaining testable implications from the general equilibrium
paradigm. Model testing serves as a barometer for measuring whether a given
parametric structure captures the essential features of the data. When cleverly
executed, it can pinpoint defective features of a model. Applied statistical
decision theory and conventional statistical practice provide a formalism for
conducting this endeavor. While this theory can be criticized for its rigidity or
its naiveté, it seems premature to scrap it altogether without putting in place
some other clearly stated criterion for picking the parameters of a model and
assessing the quality of that selection.
The rational agents in a model of the Kydland-Prescott type rely explicitly
on loss functions. After all, their rational decision making is based on the
application of statistical decision theory, and part of the Kydland-Prescott line
of research is to welcome the application of this theory to modern
macroeconomics. But the idea of a loss function is also a central concept in
statistical decision theory (LeCam and Yang, 1990). The rational agents in real
business cycle models use this theory and, as a consequence, are assumed to
process information in a highly structured way. Why should the producers of
estimates for the real business cycle models act differently?
Although Kydland and Prescott are not precise in this essay in stating how
calibration should be done in practice, there is much more specificity in
Prescott (1986a, p. 14), who writes: “The key parameters of the growth
model are the intertemporal and intratemporal elasticities of substitution. As
Lucas (1980, p. 712) emphasizes, ‘On these parameters, we have a wealth of
inexpensively available data from census and cohort information, from panel
data describing market conditions and so forth.’”
It is instructive to compare Prescott’s optimistic discussion of the ease of
using micro data to inform calibration with the candid and informative
discussion of the same issue by Shoven and Whalley (1992, p. 105), who
pioneered the application of calibration methods in general equilibrium
analysis. They write:
Typically, calibration involves only one year’s data or a single observation
represented as an average over a number of years. Because of the reliance
on a single observation, benchmark data typically does not identify a unique
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
EMPIRICAL FOUNDATIONS OF CALIBRATION
94
261
Journal of Economic Perspectives
set of values for the parameters in any model. Particular values for the
relevant elasticities are usually required, and are specified on the basis of
other research. These serve, along with the equilibrium observation, to
uniquely identify the other parameters of the model. This typically places
major reliance on literature surveys of elasticities; as many modelers have
observed in discussing their own work, it is surprising how sparse (and
sometimes contradictory) the literature is on some key elasticity values.
And, although this procedure might sound straightforward, it is often
exceedingly difficult because each study is different from every other.
What is noteworthy about this quotation is that the authors are describing a
deterministic general equilibrium model based on traditional models of factor
demand, sectoral output, product supply, labor supply and demand for final
products, which have been the focus of numerous micro empirical studies.
There have been many fewer micro empirical studies of the sectoral
components of the stochastic general equilibrium models used in real business
cycle theory. If there are few well-tested models that Shoven and Whalley can
pull off the shelf, is it plausible that the shelf is unusually rich in models
estimated assuming the relevant economic agents are operating in the more
general economic environments considered in real business cycle theory?
Shoven and Whalley (1992, p. 106) come close to acknowledging the
fundamental underidentification of general equilibrium systems from time
series data when they write:
[I]n some applied models many thousands of parameters are involved,
and to estimate simultaneously all of the model parameters using timeseries methods would require either unrealistically large numbers of
observations or overly severe identifying restrictions.… Thus far, these
problems have largely excluded complete econometric estimation of
general equilibrium systems in applied work.
Current real business cycle models often require many fewer parameters to
be calibrated, because they are highly aggregated. However, the extraction of
the required elasticities from microeconometric analyses is more problematic,
because the implicit economic environments invoked to justify
microeconometric estimation procedures seldom match the dynamic stochastic
single-agent models for which the micro estimates act as inputs. Microeconomic
studies rarely estimate models that can be directly applied to the aggregates
used in real business cycle theory. Moreover, as the specification of the real
business cycle models become richer, they will inevitably have to face up to
the same concerns that plague Shoven and Whalley.8
8
This problem has already surfaced in the work of Benhabib, Rogerson and Wright (1991).
They try to identify the parameters of a household production function for the services from
durable goods using Panel Survey of Income Dynamics data, but without data on one of the
inputs (the stock of durable goods), poor data on the other input (time spent by the household
required to make durable goods productive) and no data on the output.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
262
METHODOLOGY
The Empirical Foundations of Calibration
95
The Real Business Cycle Empirical Method In Practice
Kydland and Prescott, along with other real business cycle practitioners, endorse
the use of time series averages—but not correlations—in calibrating models. In
their proposed paradigm for empirical research, correlations are to be saved
and used to test models, but are not to be used as a source of information about
parameter values. It has become commonplace in the real business cycle research
program to match the steady-state implications of models to time series averages.
To an outsider, this looks remarkably like a way of doing estimation without
accounting for sampling error in the sample means. In fact, the real business
cycle “calibration” estimator of the Cobb-Douglas share parameter is a classical
econometric estimator due to Klein (Klein, 1953; Nerlove, 1965). The only
difference is that the Klein estimator usually is presented with a standard error.
Why is it acceptable to use sample means as a valid source of information
about model parameters and not sample autocorrelations and cross
correlations? Many interesting parameters cannot be identified from population
means alone. Although the real business cycle literature provides no good
reason for not using other sample moments, some reasons could be adduced.
For example, one traditional argument for using sample means is that they are
robust to measurement error in a way that sample variances and covariances
are not as long as the errors have mean zero. Another possible rationale is that
steady-state relations are sometimes robust with respect to alternative
specifications of the short-run dynamics of a model. In these cases, a calibration
fit to sample means will be consistent with a class of models that differ in their
implications for short-run dynamics. However, the other side of this coin is that
long-term means identify the short-run dynamics of a model only in very
special circumstances. Moreover, as pointed out by Sargent (1989), even with
measurement error, time series correlations and cross correlations can still
provide more information about a model than is conveyed in sample means.
Since the models considered by Kydland and Prescott are stochastic, it is
not in general possible to calibrate all of the parameters of a model solely
from the means of macro time series. Computational experiments make
assumptions about the correlation among the stochastic inputs to the model.
Information about shocks, such as their variances and serial correlations, are
needed to conduct the computational experiments. In a related vein,
macroeconomic correlations contain potentially valuable information about
the mechanism through which shocks are transmitted to macro time series.
For models with richer dynamics, including their original “time-to-build”
model, Kydland and Prescott (1982) envision fully calibrating the
transmission mechanisms from micro evidence; but they provide no defense
for avoiding the use of macro correlations in that task.
Recently, Cogley and Nason (1995) have criticized models in the literature
spawned by Kydland and Prescott for failing to generate business cycle
dynamics (see also Christiano, 1988; Watson, 1993; Cochrane, 1994). Since
matching the full set of dynamics of the model to the dynamics in the data is
not an essential part of calibration methodology, these models survive the
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
EMPIRICAL FOUNDATIONS OF CALIBRATION
96
263
Journal of Economic Perspectives
weak standards for verification imposed by the calibrators. A much more
disciplined and systematic exploration of the intertemporal and cross
correlations, in a manner now routine in time series econometrics, would have
shifted the focus from the empirical successes to the empirical challenges. We
agree with Oreskes, Shrader-Frechette and Belitz (1994) that the distinction
between calibration and verification is commonly blurred in practice. In the
case of real business cycle research, such blurring is likely to be all the more
prevalent as the models are redesigned to incorporate richer transient
dynamics and additional sources of uncertainty.
As Kydland and Prescott emphasize, one of the most important questions
for macroeconometrics is the quantitative importance of alternative sources
of business cycle fluctuations. This classical problem has not yet been
definitively answered (Morgan, 1990, pt. I). Using intuition from factor
analysis, it is impossible to answer this question from a single time series.
From two time series, one can isolate a single common factor. (Intuitively, two
random variables can always be decomposed into a common component and
two uncorrelated components.) Only using multiple time series is it possible
to sort out multiple sources of business cycle shocks. The current emphasis in
the literature on using only a few “key correlations” to check a model’s
implications makes single-factor explanations more likely to emerge from real
business cycle analyses.9 The idiosyncratic way Kydland and Prescott quantify
the importance of technology shocks unfortunately makes it difficult to
compare their answers to those obtained from the “innovation accounting”
methods advocated by Sims (1980) and used extensively in empirical
macroeconomics or to those obtained using the dynamic factor models of
Geweke (1977) and Sargent and Sims (1977). Kydland and Prescott’s answer
to the central question of the importance of technology shocks would be
much more credible if it were reinforced by other empirical methodologies.
A contrast with John Taylor’s approach to investigating the properties of
models is instructive. Taylor’s research program includes the use of
computational experiments. It is well summarized in his recent book (Taylor,
1993). Like Kydland and Prescott, Taylor relies on fully specified dynamic
models and imposes rational expectations when computing stochastic
equilibria. However, in fitting linear models he uses all of the information on
first and second moments available in the macro data when it is
computationally possible to do so. The econometric methods used in
parameter estimation are precisely described. Multiple sources of business
cycle shocks are admitted into the model at the outset, and rigorous empirical
testing of models appears throughout his analyses.10
9
In private correspondence, John Taylor has amplified this point: “I have found that the
omission of aggregate price or inflation data in the Kydland-Prescott second moment exercise
creates an artificial barrier between real business cycle models and monetary models. To me,
the Granger causality from inflation to output and vice versa are key facts to be explained.
But Kydland and Prescott have ignored these facts because they do not fit into their models.”
10
Fair (1994) presents an alternative systematic approach to estimation and simulation, but
unlike Taylor, he does not impose rational expectations assumptions.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
264
METHODOLOGY
Lars Peter Hansen and James J.Heckman
97
If the Kydland and Prescott real business cycle research program is to
achieve empirical credibility, it will have to provide a much more
comprehensive assessment of the successes and failures of its models. To
convince a wide audience of “outsiders,” the proclaimed successes in real
business cycle calibration should not be intertwined with an idiosyncratic and
poorly justified way of evaluating models. We sympathize with Fair (1992, p.
141), who writes:
Is the RBC [real business cycle] approach a good way of testing models?
At first glance it might seem so, since computed paths are being compared
to actual paths. But the paths are being compared in a very limited way in
contrast to the way that the Cowles Commission approach would compare
them. Take the simple RMSE [root mean square error11] procedure. This
procedure would compute a prediction error for a given variable for each
period and then calculate the RMSE from another structural model or
from an autoregressive or vector autoregressive model.
I have never seen this type of comparison done for a RBC model.
How would, say, the currently best-fitting RBC model compare to a
simple first-order autoregressive equation for real GNP in terms of the
RMSE criterion? My guess is very poorly. Having the computed path
mimic the actual path for a few selected moments is a far cry from
beating even a first-order autoregressive equation (let alone a structural
model) in terms of fitting the observations well according to the RMSE
criterion. The disturbing feature of the RBC literature is there seems to
be no interest in computing RMSEs and the like. People generally seem
to realize that the RBC models do not fit well in this sense, but they
proceed anyway.
Specification Uncertainty Underlies the Estimates
One of the most appealing features of a research program that builds dynamic
macroeconomic models on microeconomic foundations is that it opens the
door to the use of micro empirical evidence to pin down macro parameter
values. Kydland and Prescott and the entire real business cycle community
pay only lip service to the incompatibility between the macroeconomic model
used in their computational experiments and the microeconometric models
used to secure the simulation parameters.
It can be very misleading to plug microeconometric parameter estimates
into a macroeconomic model when the economic environments for the two
models are fundamentally different. In fact, many of the micro studies that
the “calibrators” draw upon do not estimate the parameters required by the
models being simulated.
11
RMSE is the square root of the mean of the squared discrepancies between predicted and
actual outcomes.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
EMPIRICAL FOUNDATIONS OF CALIBRATION
98
265
Journal of Economic Perspectives
This creates specification uncertainty (Learner, 1978). To adequately
represent this uncertainty, it is necessary to incorporate the uncertainty about
model parameters directly into the outputs of simulations. Standard errors
analogous to those presented by Christiano and Eichenbaum (1992) and
Eichenbaum (1991) are a useful first step, but do not convey the full picture
of model uncertainty. What is required is a sensitivity analysis to see how the
simulation outputs are affected by different choices of simulation parameters.
Trostel (1993) makes effective use of such a methodology.
Consider using the estimates of intertemporal labor supply produced by
Ghez and Becker (1975) for simulation purposes.12 Ghez and Becker (1975)
estimate the intertemporal substitution of leisure time assuming perfect credit
markets, no restrictions on trade in the labor market and no fixed costs of
work. This study is important, but like all empirical work in economics, the
precise estimates are enveloped by some uncertainty. Moreover, different
estimation schemes are required to secure this parameter if there is
uninsurable uncertainty in the environment (MaCurdy, 1978). Even looking
only at estimates of the intertemporal substitution of leisure based on models
that assume that workers can perfectly insure, the point estimates reported in
the literature are very imprecisely determined (MaCurdy, 1981; Altonji,
1986). Further, it is not clear how the estimates should be modified to be
compatible with the other economic environments including settings that
allow for uninsurable uncertainty, transactions costs and restrictions on trades
in the market.
Current practices in the field of calibration and simulation do not
report either estimation error and/or model-specification error. Nor is it a
standard feature of real business cycle practice to present formal analyses
that explore how sensitive the simulations are to different parameter
values. Precise numerical outputs are reported, but with no sense of the
confidence that can be placed in the estimates. This produces a false sense
of precision.
Observationally Equivalent Models Offer Different Predictions
about Policy Interventions
While putting on empirical “blinders” permits a particular line of research to
proceed, looking at too narrow of a range of data makes identification problems
more severe. A disturbing feature of current practice in the real business cycle
12
Kydland and Prescott cite Ghez and Becker (1975) as a prime example of the value of
microeconomic empirical work. However, their citation misses two central aspects of that
work. First, Ghez and Becker (1975) use synthetic cohort data, not panel data as stated by
Kydland and Prescott. Second, the interpretation of the Ghez-Becker estimates as structural
parameters is predicated on a list of identifying assumptions. These assumptions coupled
with the resulting estimates are the most important part of their investigation, not their
observation that people sleep eight hours a day, which is what Kydland and Prescott
emphasize.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
266
METHODOLOGY
The Empirical Foundations of Calibration
99
literature is that models with the same inputs can produce fundamentally
different computed values of welfare loss and quantitative assessments of
alternative economic policies.
Consider the following developments in the field of empirical finance. A
frequently noted anomaly is that the observed differences in returns between
stocks and bonds are too large to be consistent with models of the preferences
commonly used in real business cycle analysis (Hansen and Singleton, 1983;
Mehra and Prescott, 1985; Cochrane and Hansen, 1992; Kocherlakota, 1996).
One response to these asset-pricing anomalies has been the modification to
preferences developed by Epstein and Zin (1989), which breaks the tight link
between intertemporal substitution and risk aversion that was maintained in the
preceding literature. A parallel advance has been the introduction of
intertemporal complementarities such as habit persistence in the preference
orderings of consumers (Constantinides, 1990). Hansen, Sargent and Tallarini
(1995) find that models with Epstein-Zin type preferences and models without
this form of risk sensitivity explain the same quantity data but have
fundamentally different implications for the market price of risk (the slope of
the mean-standard deviation frontier for asset returns). 13 These
“observationally equivalent” preference specifications produce very different
estimates of the welfare losses associated with hypothetical policy interventions.
The decision by other researchers such as Epstein and Zin to look more broadly
at available data and to emphasize model defects instead of successes provoked
quantitatively important advances in economic theory.
Another competing explanation for the equity premium puzzle is the
presence of incomplete markets and transactions costs in asset markets. This
explanation is consistent with Prescott’s (1986b, p. 29) earlier argument for
ignoring asset market data in real business cycle calibrations: “That the
representative agent model is poorly designed to predict differences in
borrowing and lending rates…does not imply that this model is not well
suited for other purposes—for predicting the consequences of technology
shocks for fluctuations in business cycle frequencies, for example.”
Heaton and Lucas (1995) quantify the magnitude of transaction costs
needed to address the equity-premium puzzle (see also Aiyagari and Gertler,
1991). Prescott may be correct that such models will not help to match “key”
correlations in economic aggregates, but this requires documentation. Even if
there is robustness of the form hoped for by Prescott, the presence of
transactions costs of the magnitude suggested by Heaton and Lucas (1995)
are likely to alter the welfare comparisons across different policy experiments
in a quantitatively important way.14 This is so because transactions costs
prevent heterogeneous consumers from equating marginal rates of
substitution and put a wedge between marginal rates of substitution and
marginal rates of transformation.
13
Recent work by Campbell and Cochrane (1995) and Boldrin, Christiano and Fisher (1995)
suggests a similar conclusion for models with strong intertemporal complementarities.
14
This sensitivity actually occurs in the “bothersome” experiment of Gmrohorgu lu (1992)
mentioned by Kydland and Prescott.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
EMPIRICAL FOUNDATIONS OF CALIBRATION
100
267
Journal of Economic Perspectives
A More Constructive Research Program
The idea of using micro data to enrich the information in macro time series
dates back at least to the writings of Tobin (1950). A careful reading of the
literature that accompanied his suggestion reveals that his idea was inherently
controversial, especially if the micro information is based on cross-section
data, and if the behavioral equations are dynamic (Aigner and Simon, 1970;
Simon and Aigner, 1970). This issue was revived in the late 1970s and early
1980s when numerous economists attempted to estimate micro labor supply
equations to test the Lucas and Rapping (1969) intertemporal substitution
hypothesis. The hypothesis rests critically on consumer responses to expected
real discounted future wage movements relative to current wages. By providing
well-focused economic questions, the Lucas-Rapping model advanced the
development of empirical microeconomics by challenging economists to supply
answers. Numerous micro studies of labor supply were conducted with an eye
toward confirming or discontinuing their hypothesis (Altonji and Ashenfelter,
1980; MaCurdy, 1981; Ham, 1982; Altonji, 1986).
However, these studies reveal that even with large micro samples, it is not
possible to estimate the parameter of interest precisely. Measurement error in
micro data and selection problems often limit the value of the information in
the micro data. Macro time series or aggregated cross sections can sometimes
solve selection problems that are intractable in micro data (Heckman and Robb,
1985, pp. 168–169, 210–213). Different micro survey instruments produce
fundamentally different descriptions of the same phenomena (Smith, 1995).
Micro data are no panacea. Moreover, the recent movement in empirical
microeconomics away from economic models to “simple descriptive”
estimation schemes has reduced the supply of new structural parameters.
It is simply not true that there is a large shelf of micro estimates already
constructed for different economic environments that can be plugged without
modification into a new macro model. In many cases, estimators that are valid
in one economic environment are not well suited for another. Given the lessthan-idyllic state of affairs, it seems foolish to look to micro data as the
primary source for many macro parameters required to do simulation
analysis. Many crucial economic parameters—for example, the effect of
product inputs on industry supply—can only be determined by looking at
relationships among aggregates. Like it or not, time series evidence remains
essential in determining many fundamentally aggregative parameters.
A more productive research program would provide clearly formulated
theories that will stimulate more focused microeconomic empirical research.
Much recent micro research is atheoretical in character and does not link up
well with macro general equilibrium theory. For example, with rare
exceptions, micro studies treat aggregate shocks as nuisance parameters to be
eliminated by some trend or dummy variable procedure.15 A redirection of
micro empirical work toward providing input into well-defined general
15
For an exception see Heckman and Sedlacek (1985), who show how cross-section time
dummies can be used to estimate the time series of unobserved skill prices in a market model
of self-selection.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
268
METHODOLOGY
Lars Peter Hansen and James J.Heckman
101
equilibrium models would move discussions of micro evidence beyond
discussions of whether wage or price effects exist, to the intellectually more
important questions of what the micro estimates mean and how they can be
used to illuminate well-posed economic questions. “Calibrators” could make a
constructive contribution to empirical economics by suggesting a more
symbiotic relationship between the macro general equilibrium model as a
synthesizing device and motivating vehicle and the micro evidence as a source
of robust parameter values.
Recently there has been considerable interest in heterogeneous agent
models in the real business cycle literature; Ríos-Rull (1995) offers a nice
summary. To us, one of the primary reasons for pushing this line of research
is to narrow the range of specification errors in calibrating with
microeconomic data. Microeconometric estimates routinely incorporate
heterogeneity that is often abstracted from the specification of dynamic,
stochastic general equilibrium models. It is remarkable to us that so little
emphasis has been given to the transition from micro to macro in the real
business cycle literature, given that understanding the distribution of
heterogeneity is central to making this transition (Stoker, 1993).
The Kydland and Prescott program is an intellectually exciting one. To
date, however, the computations produced from it have only illustrated some
of the qualitative properties of some dynamic stochastic models and
demonstrated the possibility of executing an array of interesting calculations.
The real business cycle modeling effort would be more beneficial if it shifted
its focus to micro predictions and in this way helped to stimulate research on
empirical models that would verify or contradict the macro models.
We envision a symbiotic relationship between calibrators and empirical
economists in which calibration methods like those used by Frisch, Tinbergen,
and Kydland and Prescott stimulate the production of more convincing micro
empirical estimates by showing which gaps in our knowledge of micro
phenomenon matter and which gaps do not. Calibration should only be the
starting point of an empirical analysis of general equilibrium models. In the
absence of firmly established estimates of key parameters, sensitivity analyses
should be routine in real business cycle simulations. Properly used and
qualified simulation methods can be an important source of information and
an important stimulus to high-quality empirical economic research.
The research program we advocate is not an easy one. However, it will be
an informative one. It will motivate micro empirical researchers to focus on
economically interesting questions; it will secure the foundations of empirical
general equilibrium theory; and, properly executed, it will demonstrate both
the gaps and strengths of our knowledge on major issues of public policy.
ƒ We thank Jennifer Boobar, John Cochrane, Marty Eichenbaum, Ray Fair, Chris Flinn, John Heaton, Bob
Lucas, Tom Sargent, Jeff Smith, Nancy Stokey, John Taylor and Grace Tsiang for their valuable comments
on this draft. Hansen’s research is supported in part by NSF SBR-94095–01; Heckman is supported by
NSF 93–048–0211 and grants from the Russell Sage Foundation and the American Bar Foundation.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
EMPIRICAL FOUNDATIONS OF CALIBRATION
269
102 Journal of Economic Perspectives
References
Aigner, Dennis, and Julian L.Simon, “A
Specification BIAS Interpretation of CrossSection vs. Time-Series Parameter Estimates,”
Western Economic Journal, June 1970, 2 2,
144–61.
Aiyagari, S.Rao, and Mark Gertler,
“Asset Returns with Transitions Costs and
Uninsured Individual Risk,” Journal of Monetary Economics, June 1991, 27, 311–31.
Altonji, Joseph G., “Intertemporal Substitution in Labor Supply: Evidence from
Micro Data,” Journal of Political Economy, June
1986, 94:3, Part 2, S176–S215.
Altonji, Joseph G., and Orley
Ashenfelter, “Wage Movements and the
Labour Market Equilibrium Hypothesis,”
Economica, August 1980, 49, 783–824.
Benhabib, J., R.Rogerson, and
R.Wright, “Homework in Macroeconomics: Household Production and Aggregate
Fluctuations,” Journal of Political Economy, December 1991, 99, 1166–87.
Boldrin,
Michele,
Lawrence
J.Christiano, and J.D.M. Fisher, “Asset Pricing Lessons for Modeling Business Cycles,”
manuscript, August 1995.
Brown, D.J., and R.L.Matzkin, “Testable Restrictions on the Equilibrium Manifold,” manuscript, April 1995.
Campbell, John, and John H.Cochrane,
“By Force of Habit: A Consumption-Based
Explanation of Aggregate Stock Market Behavior.” National Bureau of Economic Research Working Paper No. 4995, 1995.
Christiano, Lawrence J., “Why Does
Inventory Investment Fluctuate So Much?,”
Journal of Monetary Economics, March/ May
1988, 21, 247–80.
Christiano, Lawrence J., and Martin
Eichenbaum, “Current Real Business Cycle
Theories and Aggregate Labor Market Fluctuations,” American Economic Review, June 1992,
82, 430–50.
Cochrane, John H., “Shocks,” CarnegieRochester Conference. Series on Public Policy,
Spring 1994, 41, 295–364.
Cochrane, John H., and Lars Peter
Hansen, “Asset Pricing Explorations for
Macroeconomics.” In Blanchard, O.J., and
S.Fischer, eds., NBER Macroeconomics Annual
1992. Cambridge, Mass.: Massachusetts Institute of Technology Press, 1992, pp. 115–
69.
Cogley, T., and J.M.Nason, “Effects of
the Hodrick-Prescott Filter on Trend and
Difference Stationary Time Series Implications for Business Cycle Research,” Journal
of Economic Dynamics and Control, January/February 1995, 19, 253–78.
Constantinides, George M., “Habit Formation: A Resolution of the Equity Premium
Puzzle,” Journal of Political Economy, June 1990,
98, 519–43.
Dunsmuir, W., and E.J.Hannan, “Vector Linear Time Series Models,” Advances in
Applied Probability, 1978, 8:2, 339–64.
Eichenbaum, M., “Real Business
Theory: Wisdom or Whimsy,” Journal of
Economic Dynamics and Control, October 1991,
15, 607–26.
Epstein, Larry G., and Stanley E.Zin,
“Substitution, Risk Aversion, and the Temporal Behavior of Consumption and Asset
Returns: A Theoretical Framework,”
Econometrica, July 1989, 57, 937–69.
Fair, Ray, “The Cowles Commission
Approach, Real Business Cycle Theories,
and New-Keynesian Economics.” In
Belongia, M.T., and M.R.Garfinkel, eds., The
Business Cycle: Theories and Evidence. Boston:
Kluwer Academic, 1992, pp. 133–47.
Fair, Ray, Testing Macroeconomic Models.
Cambridge: Harvard University Press, 1994.
Frisch, Ragnar, “Propagation Problems
and Impulse Problems in Dynamic Economics.” In Economic Essays in Honor of Gustav Cassel.
London: Allen and Unwin, 1933, pp. 171–
205.
Geweke, John, “The Dynamic Factor
Analysis of Economic Time Series Models.”
In Aigner, D.J., and A.S.Goldberger, eds.,
Latent Variables in Socio-Economic Models.
Amsterdam: North-Holland, 1977, pp. 365–
83.
Ghez, G., and Gary Becker, The Allocation of Time and Goods Over the Life Cycle. New
York: National Bureau of Economic Research,
1975.
Goldberger, Arthur, Impact Multipliers and
Dynamic Properties of the Klein-Golberger Model.
Amsterdam: North-Holland, 1959.
German, William M., “Community Preference Fields,” Econometrica, January 1953,
21, 63–80.
Ham, John C., “Estimation of a Labour
Supply Model with Censoring Due to Unemployment and Underemployment,” Review
of Economic Studies, July 1982, 49, 335–54.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
270
METHODOLOGY
The Empirical Foundations of Calibration
Hansen, Lars Peter, “Large Sample Properties of Generalized Method of Moments
Estimators,” Econometrica, July 1982, 5 0,
1029–54.
Hansen, Lars Peter, and Thomas
J.Sargent, “Seasonality and Approximation
Errors in Rational Expectations Models,”
Journal of Econometrics, January/February 1993,
55, 21–55.
Hansen, Lars Peter, and K.J.Singleton,
“Stochastic Consumption, Risk Aversion,
and the Temporal Behavior of Asset Returns,” Journal of Political Economy, 1983, 91:2,
249–65.
Hansen, Lars Peter, Thomas J.Sargent,
and T.D.Tallarini, “Pessimism and Risk in
General Equilibrium,” manuscript, 1995.
Harrison, J.Michael, and David
M.Kreps, “Martingales and Arbitrage in
Multiperiod Securities Markets,” Journal of
Economic Theory, June 1979, 20, 381–408.
Heaton, John, and D.Lucas, “Evaluating the Effects of Incomplete Markets on Risk
Sharing and Asset Pricing,” Journal of Political
Economy, forthcoming, June 1996.
Heckman, James, and Richard Robb,
Jr., “Alternative Methods for Evaluating the
Impact of Interventions.” In Heckman, J.,
and Burton Singer, eds., Longitudinal Analysis
of Labor Market Data. Cambridge, U.K.: Cambridge University Press, 1985, pp. 156–245.
Heckman, James J., and Guilherme
Sedlacek, “Heterogeneity, Aggregation, and
Market Wage Functions: An Empirical Model
of Self-Selection in the Labor Market,” Journal of Political Economy, December 1985, 93,
1077–125.
Houthakker, Hendrik, “The Pareto Distribution and the Cobb-Douglas Production
Function,” Review of Economic Studies, 1956,
23:1 27–31.
G mrohorg
u lu, Ayse, “The Welfare Costs
of Inflation Under Imperfect Insurance,” Journal of Economic Dynamics and Control, January
1992, 16, 79–91.
Klein, L.R., A Textbook of Econometrics.
Evanston, Ill.: Row, Peterson and Company,
1953.
Klein, L.A., and A.S.Goldberger, An
Econometric Model of the United States, 1929–
52. Amsterdam: North-Holland, 1955.
Kocherlakota, N., “The Equity Premium: It’s Still a Puzzle,” Journal of Economic
Literature, forthcoming 1996.
Kydland, Finn E., and Edward
C.Prescott, “Time to Build and Aggregate
103
Fluctuations,” Econometrica, 1982, 5 0:6,
1345–70.
Learner, E.E., Specification Searches: Ad Hoc
Inference with Nonexperimental Data. New York:
John Wiley & Sons, 1978.
LeCam, L., and G.Yang, Asymptotics in
Statistics: Some Basic Concepts. Berlin: SpringerVerlag, 1990.
Liu, T.C., “Underidentification, Structural
Estimation, and Forecasting,” Econometrica,
October 1960, 28, 855–65.
Lucas, Robert E., Jr., “Methods and Problems in Business Cycle Theory,” Journal of
Money, Credit and Banking, November 1980,
12, 696–715; reprinted in Lucas, R.E., ed.,
Studies in Business Cycle Theory. Cambridge,
Mass.: Massachusetts Institute of Technology
Press, 1981, pp. 271–96.
Lucas, Robert E., Jr., and Edward
C.Prescott, “Investment Under Uncertainty,”
Econometrica, 1971, 39:5, 659–81.
Lucas, Robert E., Jr., and Leonard
A.Rapping, “Real Wages, Employment, and
Inflation,” Journal of Political Economy, September/October 1969, 77, 721–54.
Lucas, Robert E., Jr., and Thomas
J.Sargent, “After Keynesian Macroeconomics,” Quarterly Review, Federal Reserve Bank of
Minneapolis, Spring 1979, 3, 1–16.
MaCurdy, Thomas, “Econometric Model
of the Labor Supply in a Life Setting,” Ph.D.
dissertation, University of Chicago, 1978.
MaCurdy, Thomas, “An Empirical
Model of Labor Supply in a Life Cycle Setting,” Journal of Political Economy , 1981, 89:6,
1059–85.
Mantel, Rolf R., “On the Characterization of Aggregate Excess Demand,” Journal of
Economic Theory, March 1974, 7, 348–55.
Mehra, Rajnish, and Edward
C.Prescott, “The Equity Premium: A Puzzle,”
Journal of Monetary Economics, March 1985,
15, 145–61.
Morgan, M.S., The History of Econometrics
Ideas. Cambridge: Cambridge University
Press, 1990.
Nerlove, M., Estimation and Identification
of Cobb-Douglas Production Functions. Chicago:
Rand McNally, 1965.
Oreskes, Naomi, Kristen ShraderFrechette, and Kenneth Belitz, “Verification,
Validation, and Confirmation of Numerical
Models in the Earth Sciences,” Science, February 4, 1994, 263, 641–46.
Peitgen, H.-O., and P.Richter, The Beauty
of Fractals. Berlin: Springer-Verlag, 1986.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
EMPIRICAL FOUNDATIONS OF CALIBRATION
271
104 Journal of Economic Perspectives
Prescott, Edward, “Theory Ahead of
Business-Cycle Measurement,” Carnegie-Rochester Conference on Public Policy, 1986a, 25, 11–
44; reprinted in Quarterly Review, Federal Reserve Bank of Minneapolis, Fall 1986, 10, 9–22.
Prescott, Edward, “Response to a Skeptic,” Quarterly Review, Federal Reserve Bank of
Minneapolis, Fall 1986b, 10, 28–33.
Ríos-Rull, José-Víctor, “Models with
Heterogeneous Agents.” In Cooley, T.F., ed.,
Frontiers of Business Cycle Research. Princeton,
N.J.: Princeton University Press, 1995, pp.
98–125.
Rubinstein, Mark, “An Aggregation
Theorem for Securities Markets,” Journal of
Financial Economics, September 1974, 1, 225–
44.
Sargent, Thomas J., “Two Models of
Measurements, and the Investment Accelerator,” Journal of Political Economics, 1989,
97:2, 251–87.
Sargent, Thomas J., and Christopher
A.Sims, “Business Cycle Modeling Without
Pretending to Have Too Much A Priori Economic Theory.” In Sims, Christopher, ed.,
New Methods in Business Cycle Research. Minneapolis: Federal Reserve Bank, 1977, pp. 45–
110.
Scarf, Herbert, and T.Hansen, The Computation of Economic Equilibria. New Haven:
Yale University Press, 1973.
Shoven, John B., and John Whalley, “A
General Equilibrium Calculation of the Effects of Differential Taxation of Income from
Capital in the U.S.,” Journal of Public Economics, November 1972, 1, 281–322.
Shoven, John B., and John Whalley,
Applying General Equilibrium. New York: Cambridge University Press, 1992.
Simon, Julian L., and Dennis J.Aigner,
“Cross Sectional Budget Studies, Aggregate
Time Series Studies and the Permanent In-
come Hypothesis,” American Economic Review,
June 1970, 60:2, 526–41.
Sims, Christopher A., “Macroeconomics
and Reality,” Econometrica, 1980, 48:1, 1–48.
Sims, Christopher A., “Rational Expectations Modeling with Seasonally Adjusted
Data,” Journal of Econometrics, January/February 1993, 55, 9–19.
Smith, J., “A Comparison of the Earnings Patterns of Two Samples of JTPA Eligibles,” unpublished paper, University of
Western Ontario, London, Canada, August
1995.
Sonnenschein, H., “Do Walres Identity
and Continuity Characterize the Class of
Community Excess Demand Functions?,”
Journal of Economic Theory, August 1973, 6,
345–54.
Stoker, T.M., “Empirical Approaches to
the Problem of Aggregation over Individuals,” Journal of Economic Literature, December
1993, 31:4, 1827–74.
Taylor, John B., Macroeconomic Policy in a
World Economy: From Econometric Design to Practical Operation. New York: W.W.Norton and
Company, 1993.
Tinbergen, J., Statistical Testing of Business
Cycle Theories. Vol. 2, Business Cycles in the
USA, 1919–1932, Geneva: League of Nations, 1939.
Tobin, James, “A Statistical Demand
Function for Food in the USA,” Journal of the
Royal Statistical Society, 1950, 113, Series A,
Part II, 113–41.
Trostel, P.A., “The Effect of Taxation on
Human Capital,” Econometrica, April 1993,
101, 327–50.
Watson, M.W., “Measures of Fit for Calibrated Models,” Journal of Political Economy,
1993, 101:6, 1011–41.
Wilson, Robert B., “The Theory of Syndicates,” Econometrica, January 1968, 36, 119–32.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
272
CHAPTER 15
Oxford Economic Papers 47 (1995), 24–44
FACTS AND ARTIFACTS: CALIBRATION AND
THE EMPIRICAL ASSESSMENT OF
REAL-BUSINESS-CYCLE MODELS
By KEVIN D.HOOVER
Department of Economics, University of California, Davis,
California 95616–8578, USA
1. Whither quantitative macroeconomics?
THE RELATIONSHIP between theory and data has been, from the beginning, a
central concern of the new-classical macroeconomics. This much is evident in the title
of Robert E.Lucas’s and Thomas J.Sargent’s landmark edited volume, Rational
Expectations and Econometric Practice (1981). With the advent of real-business-cycle models,
many new classical economists have turned to calibration methods. The new classical
macroeconomics is now divided between calibrators and estimators. But the debate is
not a parochial one, raising, as it does, issues about the relationships of models to
reality and the nature of econometrics that should be important to every school of
macroeconomic thought, indeed to all applied economics. The stake in this debate is
the future direction of quantitative macroeconomics. It is, therefore, critical to
understand the root issues.
Lucas begins the second chapter of his Models of Business Cycles with the remark:
Discussions of economic policy, if they are to be productive in any practical sense,
necessarily involve quantitative assessments of the way proposed policies are likely to
affect resource allocation and individual welfare. (Lucas 1987, p. 6; emphasis added)
This might appear to be a clarion call for econometric estimation. But appearances are
deceiving. After mentioning Sumru Altug’s (1989) estimation and rejection of the
validity of a variant of Finn E.Kydland and Edward C.Prescott’s (1982) real-businesscycle model (a model which takes up a large portion of his book), Lucas writes:
…the interesting question is surely not whether [the real-business-cycle model] can be
accepted as ‘true’ when nested within some broader class of models. Of course the
model is not ‘true’: this much is evident from the axioms on which it is constructed.
We know from the onset in an enterprise like this (I would say, in any effort in positive
economics) that what will emerge–at best–is a workable approximation that is useful in
answering a limited set of questions. (Lucas 1987, p. 45)
Lucas abandons not only truth but also the hitherto accepted standards of empirical
economics. Models that clearly do not fit the data, he argues, may nonetheless be
calibrated to provide useful quantitative guides to policy.
Calibration techniques are commonly applied to so-called ‘computable
general-equilibrium’ models. They were imported into macroeconomics as a
© Oxford University Press 1995
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
FACTS AND ARTIFACTS
K.D.HOOVER
273
25
means of quantifying real-business-cycle models, but now have a wide range
of applications. Some issues raised by calibration are common to all computable
general-equilibrium models; the concern of this paper, however, is with realbusiness-cycle models and related macroeconomic applications; and, as will
appear presently, these raise special issues. A model is calibrated when its
parameters are quantified from casual empiricism or unrelated econometric
studies or are chosen to guarantee that the model precisely mimics some
particular feature of the historical data. For example, in Kydland and Prescott
(1982), the coefficient of relative risk aversion is justified on the basis of
microeconomic studies, while the free parameters of the model are set to
force the model to match the variance of GNP without any attempt to find
the value of empirical analogues to them.
Allan W.Gregory and Gregor W.Smith (1991, p. 3) conclude that
calibration ‘…is beginning to predominate in the quantitative application of
macroeconomic models’. While indicative of the importance of the calibration
methodology, Gregory and Smith’s conclusion is too strong. Aside from the
new classical school, few macroeconomists are staunch advocates of
calibration. Within the new classical school, opinion remains divided. Even
with reference to real-business-cycle models, some practitioners have insisted
that calibration is at best a first step, which must be followed ‘…by setting
down a metric (e.g. one induced by a likelihood function) and estimating
parameters by finding values that make the metric attain a minimum’ (Gary
Hansen and Sargent 1988, p. 293).1
Sargent advocates estimation or what Kydland and Prescott (1991) call the
‘system-of-equations approach’. Estimation has been the standard approach in
macroeconometrics for over 40 years. Sargent and like-minded new classical
economists modify the standard approach only in their insistence that the
restrictions implied by dynamic-optimization models be integrated into the
estimations. The standard of empirical assessment is the usual one: how well
does the model fit the data statistically? Lucas and Kydland and Prescott reject
statistical goodness of fit as a relevant standard of assessment. The issue at
hand might then be summarized: who is right–Lucas and Kydland and
Prescott, or Sargent?
The answer to this question is not transparent. Estimation is the status quo.
And, although enthusiastic advocates of calibration already announce its
triumph, its methodological foundations remain largely unarticulated. An
uncharitable interpretation of the calibration methodology might be that the
advocates of real-business-cycle models are so enamored of their creations
that they would prefer to abandon commonly accepted, neutral standards of
empirical evaluation (i.e. econometric hypothesis testing) to preserve their
1
Despite the joint authorship of the last quotation, I regard Sargent and not Hansen as the
preeminent proponent of the necessity of estimation, because I recall him forcefully insisting
on it in his role as discussant of a paper by Thomas F.Cooley and Hansen (1989) at the
Federal Reserve Bank of San Francisco’s Fall Acacemic Conference; see also Manuelli and
Sargent (1988, pp. 531–4).
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
274
METHODOLOGY
26
FACTS AND ARTIFACTS
models. This would be an ad hoc defensive move typical of a degenerating
research program.
This interpretation is not only uncharitable, it is wrong. Presently, we shall
see that Herbert Simon’s (1969) Sciences of the Artifical provides the materials
from which to construct a methodological foundation for calibration, and that
calibration is compatible with a well-established approach to econometrics that
is nonetheless very different from the Cowles Commission emphasis on the
estimation of systems of equations. Before addressing these issues, however,
it will be useful to describe the calibration methodology and its place in the
history and practice of econometrics in more detail.
2. The calibration methodology
2.1. The paradigm case
Kydland and Prescott (1982) is the paradigm new-classical equilibrium, realbusiness-cycle model. It is neoclassical optimal-growth model with stochastic
shocks to technology which cause the equilibrium growth path to fluctuate about
its steady state.2 Concrete functional forms are chosen to capture some general
features of business cycles. Production is governed by a constant-elasticity-ofsubstitution production function in which inventories, fixed capital, and labor
combine to generate a single homogeneous output that may either be consumed
or reinvested. Fixed capital requires a finite time to be built before it becomes a
useful input. The constant-relative-risk-aversion utility function is rigged to possess
a high degree of intertemporal substitutability of leisure. Shocks to technology
are serially correlated. Together the structure of the serial correlation of the
technology shocks and the degree of intertemporal substitution in consumption
and leisure choices govern the manner in which shocks are propagagated through
time and the speed of convergence back towards the steady state.
Once the model is specified, the next step is to parameterize its concrete
functional forms. Most of the parameters of the model are chosen from values
culled from other applied econometric literatures or from general facts about
national-income accounting. For example, Thomas Mayer (1960) estimated the
average time to construct complete facilities to be 21 months; Robert E.Hall
(1977) estimated the average time from start of projects to beginning of
production to be two years. Citing these papers, but noting that consumer
durable goods take considerably less time to produce, Kydland and Prescott
(1982, p. 1361) assume that the paramters governing capital formation are set
to imply steady construction over four quarters.3 The values for depreciation
rates and the capital/inventory ratio are set to rough averages from the national2
For general descriptions of the details and varieties of real-business-cycle models see
Lucas (1987), Kevin D.Hoover (1988, ch. 3), and Bennett T.McCallum (1989). Steven
M.Sheffrin (1989, ch. 3, especially pp. 80, 81), gives a step-by-step recipe for constructing a
prototypical real-business-cycle model.
3
Mayer’s estimates were for complete projects only, so that new equipment installed in
old plants, which must have a much shorter time-to-build than 21 months, was not counted.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
FACTS AND ARTIFACTS
K.D.HOOVER
275
27
income accounts. Ready estimates from similar sources were not available for
the remaining six paramters of the model, which include parameters governing
intertemporal substitution of leisure and the shocks to technology. These were
chosen by searching over possible parameter values for a combination that best
reproduced certain key variances and covariances of the data. In particular, the
technology shock variance was chosen in order to exactly match the variance of
output in the postwar US economy.
To test the model’s performance, Kydland and Prescott generate a large
number of realizations of the technology shocks for 118 periods
corresponding to their postwar data. They then compute the variances and
covariances implied by the model for a number of important variables:
output, consumption, investment, inventories, the capital stock, hours
worked, productivity, and the real rate of interest.4 These are then compared
with the corresponding variances and covariances of the actual US data.5
Kydland and Prescott offer no formal measure of the success of their model.
They do note that hours are insufficiently variable with respect to the variability
of productivity to correspond accurately to the data, but otherwise they are
pleased with the model’s ability to mimic the second moments of the data.
Real-business-cycle models, treated in the manner of Kydland and Prescott,
are a species of the genus computable (or applied) general-equilibrium
models. The accepted standards for implementing computable generalequilibrium models, as codified, for example, in Ahsan Mansur and John
Whalley (1984), do not appear to have been adopted in the real-businesscycle literature. For example, while some practitioners of computable generalequilibrium models engage in extensive searches of the literature in order to
get some measure of the central tendency of assumed elasticities, Kydland and
Prescott’s (1982) choice of parameterization appears almost casual. Similarly,
4
In fact, it is not clear in Kydland and Prescott (1982) that these are calculated from the
cross-section of a set of realizations or from a single time-series realization. In a subsequent
paper that extends their results, Kydland and Prescott (1988, p. 353) are quite precise about
using a cross-section of many realizations. Because they are interested only in the moments of
the variables and not in particular time-series, Kydland and Prescott initialize variables to
their steady-state values or, equivalently in the context of detrended data, to zero. In order to
generate a time path that can be compared to the history of a particular series, it is necessary,
as in Hansen and Prescott (1993), to initialize at some actual historical benchmark.
5
Despite my referring to Kydland and Prescott’s model as a growth model, the model for
which they calculate the variances and covariances does not possess an exogenous source of
trend growth. Thus, to make comparisons, Kydland and Prescott (1982, p. 1362) detrend the
actual data using the Hodrick-Prescott filter. The particular choice of filter is not defended in
any detail. Prescott (1983, p. 6) simply asserts that it produces ‘about the right degree of
smoothness, when fit to the logarithm of the real GNP series’ without any indication by what
standard rightness is to be judged. Kydland and Prescott (1990, p. 9) claim that it generates a
trend close to the trend that students of business cycles would draw by hand through a plot
of actual GNP. Although the Hodrick-Prescott filter is almost universally adopted in comparing real-business-cycle models to actual data, Fabio Canova (1991b) shows that the use of
Hodrick-Prescott filters with different choices for the values of a key parameter or of several
entirely different alternative filters radically alters the cyclical characteristics of economic
data (also see Timothy Cogley and James Nason, 1993).
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
276
28
METHODOLOGY
FACTS AND ARTIFACTS
although Kydland and Prescott report some checks on robustness, these
appear to be perfunctory.6
In the context of computable general-equilibrium models, calibration is
preferred in those cases in which, because of extensive disaggregation, the
number of parameters is too large relative to the available data set to permit
econometric estimation.7 Since typical real-business-cycle models are one-good,
one-agent models, there is no difficulty in estimating them using standard
methods such as maximum likelihood or generalized method of moments.
Indeed, since the practitioners are often interested principally in matching
selected second moments, method-of-moments estimators can concentrate on
the moments of interest to the exclusion of others (see Watson 1993, p. 1037).
As noted earlier, Altug (1989) estimates and rejects a close relative of
Kydland and Prescott’s model using maximum likelihood. The central
problem of this paper can be restated: is there a case for ignoring Altug’s
rejection of the Kydland and Prescott model? The case must be something
other than the standard one of too many parameters found in the literature
on computable general-equilibrium models.
2.2. Calibration as estimation
Various authors have attempted to tame calibration and return it to the
traditional econometric fold. Manuelli and Sargent (1988), Gregory and Smith
(1990a), Canova (1991a), and Bansal et al. (1991) interpret calibration as a
form of ‘estimation by simulation’. In such a procedure, parameters are chosen,
and the relevant features of the simulated output of the calibrated model are
compared to the analogous features of the actual data. Such a procedure
differs from standard estimation methods principally in that it allows the
investigator to expand or restrict the range of features considered to be relevant.
Lucas’s argument, however, is that any form of estimation is irrelevant. In
their early writings, Kydland and Prescott were not in fact as explicit as Lucas
about the irrelevance of estimation. They merely argued that it would be
premature to apply techniques to their model, such as those developed by Lars
Peter Hansen and Sargent (1980), to account for the systemic effects of rational
expectations (Kydland and Prescott, 1982, p. 1369). Prescott (1983, pp. 8–11)
was more pointed: real-business-cycle models are tightly parameterized. They
6
Canova (1991a) suggests a formal methodology and provides an example in which
sensitivity analysis is conducted with respect to distributions for parameter values constructed from the different values reported in unrelated studies or from a priori information on
the practically or theoretically admissible range of parameter values.
7
Lawrence J.Lau (1984) notices that any model that can be calibrated can also be estimated. He uses ‘calibration’, however, in a narrow sense. A model is calibrated when its
parameters are chosen to reproduce the data of a benchmark period. Thus, parameterization
on the basis of unrelated econometric studies does not count as calibration for him. Lau’s
usage is diametrically opposed to that of Gregory and Smith (1991) for whom calibration is
only the assignment of parameter values from unrelated sources. We use ‘calibration’ in both
Lau’s and Gregory and Smith’s senses. Lau and, similarly, James MacKinnon (1984) make
strong pleas for estimation instead of, or in addition to, calibration, and for subjecting
computable general-equilibrium models to statistical specification tests.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
FACTS AND ARTIFACTS
277
K.D.HOOVER
29
will almost inevitably be rejected against a weakly restricted alternative
hypothesis, but such alternative hypotheses arise from the introduction of
arbitrary stochastic processes and, so, are not suitable benchmarks for
economic inference. ‘A model may mimic the cycle well but not perfectly’
(Prescott 1983, p. 10). Similarly, Kydland and Prescott (1991, p. 174) write:
Unlike the system-of-equations approach, the model economy which
better fits the data is not [necessarily?] the one used. Rather, currently
established theory dictates which one is used.
The dominance of theory in the choice of models lies at the heart of the
difference between estimators and calibrators. To throw the difference into
high relief, one can think of estimators pursuing a competitive strategy and
calibrators pursuing an adaptive strategy. Under the competitive strategy,
theory proposes, estimation and testing disposes. In fine, alternative theories
compete with one another for the support of the data. The adaptive strategy
begins with an unrealistic model, in the sense of one that is an idealized and
simplified product of the core theory. It sees how much mileage it can get out
of that model. Only then does it add any complicating and more realistic
feature. Unlike the competitive strategy, the aim is never to test and possibly
reject the core theory, but to construct models that reproduce the economy
more and more closely within the strict limits of the basic theory.
The distinction between the competitive and adaptive strategies is sharply
drawn and somewhat stylized, but focuses nonetheless on a genuine
difference. On the one hand, the competitive strategy is the received view of
econometricians, taught in an idealized form in most econometric textbooks,
even if more honored in the breach than the observance by applied
economists. The competitive strategy is explicit in Gregory and Smith’s
(1990b) ‘Calibration as Testing’. Even if in practice no theory is ever decisively
rejected through a test based on an econometic estimation, the theory is
nonetheless regarded as at risk and contingent—even at its core. On the other
hand, the real-business-cycle modeller typically does not regard the core
theory at risk in principle. Like the estimators, the calibrators wish to have a
close fit between their quantified models and the actual data—at least in
selected dimensions. But the failure to obtain a close fit (statistical rejection)
does not provide grounds for rejecting the fundamental underlying theory.
Adaptation in the face of recalcitrant data is adaptation of peripheral
assumptions, not of the core. Thus, the inability of Kydland and Prescott’s
(1982) original real-business-cycle model to match the data prompted more
complicated versions of essentially the same model that included, for
example, heterogeneous labor (Kydland 1984), a banking sector (Robert
G.King and Charles I.Plosser 1984), indivisible labor (Gary Hansen 1985),
separate scales for straight-time and overtime work (Gary Hansen and
Sargent 1988), and variable capital intensity (Kydland and Prescott 1988).
One consequence of these strategies is that esimators possess a common
ground, the performance of each theoretically-based specification against actual
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
278
METHODOLOGY
30
FACTS AND ARTIFACTS
data, on which to judge the performance of competing models. For the
calibrators, however, data help discriminate only b etween different
adaptations of the common core. The core theory itself is not questioned, so
that, unintentially perhaps, the core theory becomes effectively a Lakatosian
hardcore (Lakatos 1970, 1978; Blaug 1992, ch. 2). Calibration does not
provide a method that could in principle decide between fundamentally
different business-cycle models (e.g. real-business-cycle models or Keynesian
business-cycle models) on the basis of empirical evidence derived from the
calibration exercise itself.8 Critics of real-business-cycle models who attempt
such comparisons fall back either on attacking the discriminating power of
calibration methods (e.g. Hartley et al. 1993) or on adaptations of standard
econometric techniques (e.g. Canova et al. 1993). Kydland and Prescott are
explicit in rejecting these applications of estimation techniques as missing the
point of the calibration method. The aim of this paper is partly to explicate
and appraise their view.9
2.3. The mantel of Frisch
Calibrators radically reject the system-of-equations approach. But Kydland and
Prescott, at least, do not reject econometrics. Rather, they argue that econometrics
is not coextensive with estimation; calibration is econometrics. Kydland and
Prescott (1991, pp. 161, 162) point out that for Ragnar Frisch, Irving Fisher,
and Joseph Shumpeter, the founders of the Econometric Society, ‘econometrics’
was the unification of statistics, economic theory, and mathematics.
Unacknowledged by Kydland and Prescott, Mary Morgan’s (1990) brilliant
history of econometrics supports and elaborates their point. According to Morgan,
even before the term ‘econometrics’ had wide currency, the econometric ideal
had been to weld mathematical, deductive economics to statistical, empirical
economics to provide a substitute for the experimental methods of the natural
sciences appropriate to the study of society. This ideal collapsed with the rise of
the system-of-equations approach in the wake of the Cowles Commission.
Kydland and Prescott point to Frisch’s (1933) article, ‘Propagation
Problems and Impulse Response Problems in Dynamic Economics’ as a
precursor to both their own real-business-cycle model and to calibration
methods. Frisch argues that quantitative analysis requires complete models:
i.e. general-equilibrium models in a broad sense. He considers a sequence of
models, starting with a very simple one, and then adding complications. He
models the time-to-build feature of capital formation. He distinguishes
between the impulses that start business cycles and the dynamic
mechansisms that amplify and propagate them. He quantifies his models
using calibration techniques. And, precisely like Kydland and Prescott
8
This is not to say that there could not be some other basis for some decision.
Hoover 1994a (as well as work in progress) outlines a possible method of using econometric techniques in a way that respects the idealized nature of the core models without giving
up the possibility of empirical discrimination.
9
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
FACTS AND ARTIFACTS
279
K.D.HOOVER
31
(1982), Frisch marvels at how well a very simple model can capture the
features of actual data.
Although Kydland and Prescott are correct to see the affinities between Frisch’s
work and their own, they ignore the very real differences between Frisch and
themselves. Frisch’s approach is wholly macroeconomic. Frisch writes:
In order to attack these problems on a macro-dynamic basis so as to
explain the movement of the system taken in its entirety, we must
deliberately disregard a considerable amount of the details of the picture.
We may perhaps start by throwing all kinds of production into one
variable, all consumption into another, and so on, imagining that the
notions ‘production’, ‘consumption’, and on, can be measured by some
sort of total indices. (1933, p. 173)
While his flight to aggregates parallels the practice of the new classical realbusiness-cycle model, Frisch does not suggest that this is a way station on the
road to microfoundations. His article does not hint at the desirability of
microfoundations, even of the pseudo-microfoundations of the representativeagent model: there is not an optimization problem to be found. Frisch appears
to use calibration mainly for purposes of illustration, and not to advocate it as
a preferred technique. He writes:
At present I am guessing very roughly at these parameters, but I believe
that it will be possible by appropriate statistical methods to obtain more
exact information about them. I think, indeed, that the statistical
determination of such structural parameters will be one of the main
objectives of the economic cycle analysis of the future. (1933, p. 185)
Precisely which statistical methods are appropriate appears to be an open
question.10
More generally, although Frisch stresses the importance of theory, there is
no hint that his interpretation is limited to ‘maximizing behavior subject to
constraints’ (Kydland and Prescott 1991, p. 164). Frisch does not define
‘theory’ in ‘Propagation Problems…’, but the examples he produces of
theories are not of an obviously different character from the structures
employed by Jan Tinbergen, Lawrence Klein, James Dusenberry, and the
other ‘Keynesian’ macromodelers who are the special bugbears of the
advocates of new-classical, real-business-cycle models.
Schumpeter (co-founder with Frisch of the Econometric Society) provides
typically prolix discussions of the meaning of ‘economic theory’ in his magisteral
History of Economic Analysis (1954). For Schumpeter (1954, pp. 14, 15), theories
10
Frisch’s own shifting views illustrate how open a question this was for him. By 1936, he
had backtracked on the desirability of estimating calibrated models. In 1938, he argued that
structural estimation was impossible because of pervasive multicollinearity. In its place he
proposed estimating unrestricted reduced forms (see Morgan, 1990, p. 97; also see Aldrich
1989, section 2). In this he comes much closer to Christopher Sims (1980) program of vector
autoregressions without ‘incredible’ identifying restrictions. I am grateful to an anonymous
referee for reminding me of this point. Hoover (1992) identifies Sims’s program as one of
three responses to the Lucas (1976) critique of policy invariance. Calibration and the systems-of-equations approach each possess a analogous response.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
280
32
METHODOLOGY
FACTS AND ARTIFACTS
are, on the one hand, ‘synonomous with Explanatory Hypotheses’, and on the
other hand, ‘the sum total of the gadgets’, such as ‘“marginal rate of
substitution”, “marginal productivity”, “multiplier”, “accelerator”’, including
‘stragegically useful assumptions’, ‘by which results may be extracted from the
hypothesis’. Schumpeter concludes: ‘In Mrs. Robinson’s unsurpassingly
felicitous phrase, economic theory is a box of tools’. Later Schumpeter defends
the theoretical credentials of Wesley C.Mitchell, the subject of Tjalling
Koopman’s (1947) famous attack on ‘Measurement without Theory’:
…in intention as well as in fact, he was laying the foundations for a
‘theory’, a business cycle theory as well as a general theory of the economic
process, but for a different one. (1954, p. 1166)
Kydland and Prescott (1991, p. 164) argue that the system-of-equations
approach flourished in the 1950s only because economists lacked the tools to
construct stochastic computable general-equilibrium models. They proclaim
the death of the system-of-equations approach:
The principal reason for the abandonment of the system-of-equations
approach, however, was the advances in neoclassical theory that permitted
the application of the paradigm in dynamic stochastic settings. Once the
neoclassical tools needed for modeling business cycle fluctuations existed,
their application to this problem and their ultimate domination over any
other method was inevitable. (1991, p. 167)
This is an excessively triumphalist and whiggish history of the development of
econometric thought. First, the work of Frisch and others in the 1930s provides
no support for Kydland and Prescott’s narrowing of the meaning of ‘theory’ to
support such tendentious statements as: ‘To summarize the Frisch view, then,
econometrics is quantitative neoclassical theory with a basis in facts’ (Kydland and
Prescott 1991, p. 162; emphasis added). (A model is ‘neoclassical’ for Kydland
and Prescott (1991, p. 164) when it is constructed from ‘…agents maximizing
subject to constraints and market clearing’.) Second, the declaration of the death
of the system-of-equations approach is premature and greatly exaggerated.
Allegiance to the system-of-equations approach motivates the many efforts to
interpret calibration as a form of estimation. Third, the calibration methodology
is not logically connected to Kydland and Prescott’s preferred theoretical
framework. The example of Frisch shows that calibration can be applied to
models that are not stochastic dynamic optimal-growth models. The example of
Lars Peter Hansen and Sargent (1980) shows that, even those who prefer such
models, can use them as the source of identification for systems of equations—
refining rather than supplanting the traditional econometrics of estimation.
3. The quantification of theory
Although Kydland and Prescott overstate the degree to which Frisch and the
econometrics of the 1930s foreshadowed their work, they are nonetheless correct
to note many affinities. But such affinities, even if they were more complete
than they turn out to be, do not amount to an argument favoring calibration
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
FACTS AND ARTIFACTS
281
K.D.HOOVER
33
over estimation. At most, they are an illicit appeal to authority. To date, no
compelling defence of the calibration methodology has been offered. An
interpretation of the point of calibration and an assessment of its merits can
be constructed, however, from hints provided in Lucas’s methodological writings
of the 1970s and early 1980s.
3.1. Models
‘Model’ is a ubiquitous term in economics, and a term with a variety of meanings.
One commonly speaks of an econometric model. Here one means the concrete
specification of functional forms for estimation. I call these observational models.
The second main class of models are evaluative or interpretive models. An
obvious subclass of interpretive/evaluative models are toy models.
A toy model exists merely to illustrate or to check the coherence of
principles or their interaction. An example of a toy model is the overlappinggenerations model with money in its simplest incarnations. No one would
think of drawing quantitative conclusions about the working of the economy
from it. Instead one wants to show that models constructed on its principles
reproduce certain known qualitative features of the economy and suggest
other qualitative features that may not have been known or sufficiently
appreciated (cf. Diamond 1984, p. 47), Were one so rash as to estimate such
a model, it would surely be rejected, but that would be no reason to abandon
it as a testbed for general principles.
Is there another subclass of interpretive/evaluative models, one that
involves quantification? Lucas seems to think so:
One of the functions of theoretical economics is to provide fully articulated, artificial
economic systems that can serve as laboratories in which policies that would be
prohibitively expensive to experiment within actual economies can be tested out at
much lower cost. (Lucas 1980, p. 271)
Let us call such models benchmark models. Benchmark models must be abstract
enough and precise enough to permit incontrovertible answers to the questions
put to them. Therefore,
…insistence on the ‘realism’ of an economic model subverts its potential usefulness in
thinking about reality. Any model that is well enough articulated to give clear answers
to the questions we put to it will necessarily be artificial, abstract, patently unreal.
(Lucas 1980, p. 271)
On the other hand, only models that mimic reality in important respects will
be useful in analyzing actual economies.
The more dimensions in which the model mimics the answers actual economies give to
simple questions, the more we trust its answers to harder questions. This is the sense in
which more ‘realism’ in a model is clearly preferred to less. (Lucas 1980, p. 272)
Later in the same essay, Lucas emphasizes the quantitative nature of such
model building:
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
282
34
METHODOLOGY
FACTS AND ARTIFACTS
Our task…is to write a FORTRAN program that will accept specific economic
policy rules as ‘input’ and will generate as ‘output’ statistics describing the
operating characteristics of time series we care about, which are predicted
to result from these policies, (p. 288)
For Lucas, Kydland and Prescott’s model is precisely such a program.11
One might interpret Lucas’s remarks as making a superficial contribution
to the debate over Milton Friedman’s ‘Methodology of Positive Economics’
(1953): must the assumptions on which a theory is constructed be true or
realistic or is it enough that the theory predicts ‘as if’ they were true? But
this would be a mistake. Lucas is making a point about the architecture of
models and not about the foundations of secure prediction. Lucas refers to
a model as fully ‘realistic’ when it fully accounts for all the factors that
determine the variables of interest. Lucas makes two points.
Uncontroversially, he argues that toy models convey deep understanding of
economic principles. More interestingly, he argues that benchmark models
have an advantage over estimation. This is controversial because estimators
believe that fully articulated specifications are required for accurate
quantification. This is expressed in their concern for specification error,
omitted variable bias, and so forth. Their view is widely shared. The point is
not that estimated models are necessarily more realistic in Lucas’s sense
than calibrated models, nor that estimation is the only or even the most
reliable way to quantify a model or its components.12 Rather it is that any
method of quantification that does not aspire to full articulation is likely to
mislead. Lucas denies this, and the interesting issues are how to appraise his
position, and, if his position is sustainable, how to appraise quantified
benchmark models themselves.
To make this clear, consider Lucas’s (1987, pp. 20–31) cost-benefit analysis of
the policies to raise GNP growth and to damp the business cycle. Lucas’s model
considers a single representative consumer with a constant-relative-risk-aversion
utility function facing an exogenous consumption stream. The model is
calibrated by picking reasonable values for the mean and variance of
consumption, the subjective rate of discount, and the constant coefficient of
relative risk aversion. Lucas then calculates how much extra consumption
consumers would require to compensate them in terms of utility for a cut in the
growth of consumption and how much consumption they would be willing to
give up to secure smoother consumption streams. Although the answers that
Lucas seeks are quantitative, the model is not used to make predictions that
might be subjected to statistical tests. Indeed, it is a striking illustration of why
calibration should not be interpreted as estimation by simulation. Lucas’s model
is used to set upper bounds to the benefits that might conceivably be gained in
the real world. Its parameters must reflect some truth about the world if it is to
11
Kydland and Prescott do not say, however, whether it is actually written in FORTRAN.
For example, to clarify a point raised by an anonymous referee, if the central bank had
direct knowledge of it money supply function, that would be better than estimating it.
12
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
FACTS AND ARTIFACTS
283
K.D.HOOVER
35
be useful, but they could not be easily directly estimated. In that sense, the
model is unrealistic.13
3.2. Artifacts
In a footnote, Lucas (1980, p. 272, fn. 1) cites Simon’s Sciences of the Artificial
(1969) as an ‘immediate ancestor’ of his ‘condensed’ account. To uncover a
more fully articulated argument for Lucas’s approach to modelling, it is worth
following up the reference.
For Simon, human artifacts, among which he must count economic models,
can be thought of as a meeting point—an ‘interface’…—between an ‘inner’ environment,
the substance and organization of the artifact itself, and an ‘outer’ environment, the
surroundings in which it operates. (Simon 1969, pp. 6, 7)
An artifact is useful, it achieves its goals, if its inner environment is appropriate
to its outer environment.
The distinction between the outer and inner environments is important
because there is some degree of independence between them. Clocks tell time
for the outer environment. Although they may indicate the time in precisely
the same way, say with identical hands on identical faces, the mechanisms of
different clocks, their inner environments, may be constructed very
differently. For determining when to leave to catch a plane, such differences
are irrelevant. Equally, the inner environments may be isolated from all but a
few key features of the outer environment. Only light entering through the
lens for the short time that its shutter is open impinges on the inner
environment of the camera. The remaining light is screened out by the
opaque body of the camera, which is an essential part of its design.
Simon factors adaptive systems into goals, outer environments and inner
environments. The relative independence of the outer and inner
environments means that
[w]e might hope to characterize the main properties of the system and its behavior
without elaborating the detail of either the outer or the inner environments. We might
look toward a science of the artificial that would depend on the relative simplicity of
the interface as its primary source of abstraction and generality. (Simon 1969, p. 9)
Simon’s views reinforce Lucas’s discussion of models. A model is useful only
if it foregoes descriptive realism and selects limited features of reality to
reproduce. The assumption upon which the model is based do not matter, so
long as the model succeeds in reproducing the selected features. Friedman’s
‘as if’ methodology appears vindicated.
13
Of course, Lucas’s approach might be accepted in principle and still rejected in detail.
For example, McCallum (1986, pp. 411, 412) objects to the characterization of consumption
as fluctuating symmetrically about trend that is implicit in Lucas’s use of a mean/variance
model. If the fluctuations of consumption are better described as varying degrees of shortfall
relative to the trend of potential maximum consumption, then the benefits of consumption
smoothing will be considerably higher than Lucas’s finds.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
284
36
METHODOLOGY
FACTS AND ARTIFACTS
But this is to move too fast. The inner environment is only relatively independent
of the outer environment. Adaptation has its limits.
In a benign environment we would learn from the motor only what it
had been called upon to do; in a taxing environment we would learn
something about its internal structure—specifically, about those aspects
of the internal structure that were chiefly instrumental in limiting
performance. (Simon 1969, p. 13)14
This is a more general statement of principles underlying Lucas’s (1976) critique
of macroeconometric models. A benign outer environment for econometric
models is one in which policy does not change. Changes of policy produce
structural breaks in estimated equations: disintegration of the inner environment
of the models. Economic models must be constructed like a ship’s chronometer,
insulated from the outer environment so that ‘…it reacts to the pitching of the
ship only in the negative sense of maintaining an invariant relation of the hands
on its dial to real time, independently of the ship’s motions’ (Simon 1969, p. 9).
Insulation in economic models is achieved by specifying functions whose
parameters are invariant to policy. The independence of the inner and outer
environments is not something which is true of arbitrary models; rather it must
be built into models. While it may be enough in hostile environments for models
to reproduce key features of the outer environment ‘as if’ reality was described
by their inner environments, it is not enough if they can do this only in benign
environments. Thus, for Lucas, the ‘as if’ methodology interpreted as an excuse
for complacency with respect to modeling assumptions must be rejected. Simon’s
notion of the artifact helps justify Lucas’s both rejecting realism in the sense of
full articulation and at the same time, insisting that only through carefully
constructing the model from invariants—tastes and technology, in Lucas’s usual
phrase—can the model secure the benefits of a useful abstraction and generality.
Recognizing that a model must be constructed from invariants does not
itself tell us how to quantify it. The emphasis on a maintained theory or inner
environment presents a generic risk for quantified idealized models (see
Section 2.2 above). The risk is particularly severe for the calibration
methodology with its adaptive strategy. Gregory and Smith (1991, p. 30)
observe that ‘[s]etting parameter values (i.e. calibrating), simulating a model
and comparing properties of simulations to those of data often suggests
fruitful modifications of the model’. Generally, such modifications leave the
essential core theory intact and attempt to better account for the divergences
from the ideal, to better account for the fudge factors need to link the output
of the model to the phenomenal laws. The risk, then, is that the core of the
model becomes completely insulated from empirical confirmation or
disconfirmation—even in the weakest senses of those terms. Kydland and
14
Haavelmo (1944, p. 28) makes a similar point in his well-known example of the failure
of autonomy: the relationship between the speed of a car and the amount of throttle may be
well-defined under uniform conditions, but would break down immediately the car was
placed in a different setting. To understand how the car will perform on the track as well as on
the road requires us to repair to the deeper principles of its operation.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
FACTS AND ARTIFACTS
K.D.HOOVER
285
37
Prescott (1991, p. 171) explicitly deny that the confidence in the answers a
model gives to policy questions can ‘…be resolved by computing some
measure of how well the model economy mimics historical data’. Rather, ’[t]he
degree of confidence in the answer depends on the confidence that is placed in
the economic theory being used’. Kydland and Prescott do not explain what
alternative sources there might be to justify our confidence in theory; the
adaptive strategy of the calibration approach almost guarantees that empirical
evidence will not be among those sources.
3.3. Quantification without history
Calibrators of real-business-cycle models typically concentrate on matching
selected second moments of variables rather than, say, matching the actual
historical evolution of the modeled variables. Why? Lucas (1977, p. 218)
observes that ‘business cycles are all alike’, not in exact detail but qualitatively.
An informative test of a model’s ability to capture business-cycle behavior is
not, therefore, its fit to some historical time series, which is but one of many
possible realizations, but rather its ability to characterize the distribution from
which that realization was drawn. Lucas (1977, pp. 219, 234) advocates the
test of Irma Adelman and Frank L.Adelman (1959). The Adelmans asked the
question, could one distinguish data generated by simulating a model (in their
case, the Klein-Goldberger macroeconometric model) from actual data
describing the economy, in the absence of knowledge of which was which?
The Adelmans’ test compares the distribution of outcomes of the model to
the actual economy. Once a close relation is established, to experiment with
alternative policy rules is an easy next step. Even though government is not
modelled in Kydland and Prescott’s initial models, policy analysis is their ultimate
goal (Kydland and Prescott, 1982, p. 1369). Concentration on the second
moments of variables can be seen as the practical implementation of the
Adelmans’ standard: one eschews the particular realization in favor of a more
general characterization of the distribution of possible outcomes.15
One reason, therefore, not to apply a neutral statistical test for the match
between model and reality is that it is along only selected dimensions that one
cares about the model’s performance at all. This is completely consistent with
Simon’s account of artifacts. New classical economics has traditionally been
skeptical about discretionary economic policies. New classical economists are,
therefore, more concerned to evaluate the operating characteristics of policy
rules. For this, the fit of the model to a particular historical realization is largely
irrelevant, unless it assures it will also characterize the future distribution of
15
The Adelmans themselves examine the time-series properties of a single draw, rather
than the characteristics of repeated draws. This probably reflects, in part, the computational
expense of simulating a large macroeconometric model with the technology of 1959. It also
reflects the application of Burns and Mitchell’s techniques for characterizing the repetitive
features of business cycles through averaging over historical cycles all normalized to a
notional cycle length. King and Plosser (1989) attempt to apply precisely Burns and Michell’s
techniques to outcomes generated by a real-business-cycle model.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
286
38
METHODOLOGY
FACTS AND ARTIFACTS
outcomes. The implicit claim of most econometrics is that it does assure a
good characterization. Probably most econometricians would reject
calibration methods as coming nowhere close to providing such assurance.
Substantial work remains to be done in establishing objective, comparative
standards for judging competing models.
4. Aggregation and general equilibrium16
Whether calibrated or estimated, real-business-cycle models are idealizations
along many dimensions. The most important dimension of idealization is the
the models deal in aggregates while the economy is composed of individuals.
After all, the distinction between microeconomics and macroeconomics is the
distinction between the individual actor and the economy as a whole. All new
classical economists believe that one understands macroeconomic behavior
only as an outcome of individual rationality. Lucas (1987, p. 57) comes close to
adopting the Verstehen approach of the Austrians.17 The difficulty with this
approach is that there are millions of people in the economy and it is not
practical—nor is it ever likely to become practical—to model the behavior of
each of them.18 Universally, new classical economists adopt representativeagent models, in which one agent or a few types of agents, stand in for the
behavior of all agents.19 The conditions under which a single agent’s behavior
can accurately represent the behavior of an entire class are onerous and almost
certainly never fulfilled in an actual economy.
One interpretation of the use of calibration methods in macroeconomics is that
the practitioners recognize that highly aggregated theoretical models must be
descriptively false, so that estimates of them are bound to fit badly in comparison
to atheoretical econometric models, which are able to exploit large numbers of
free parameters. The theoretical models are nonetheless to be preferred because
policy evaluation is possible only within their structure. In this, they are exactly
like Lucas’s benchmark consumption model (see Section 3.1, above).
Calibractors appeal to microeconomic estimates of key parameters because
information about individual agents is lost in the aggregatation process.
Estimators, in contrast, could argue that the idealized representative-agent
16
Aggregation and the problems it poses for macroeconomics are the subject of a voluminous literature. The present discussion is limited to a narrow set of issues most relevant to the
question of idealization.
17
For a full discussion of the relationship between new classical and Austrian economics
see Hoover (1988, ch. 10).
18
In Hoover (1984, pp. 64–6; and 1988, pp. 218–20), I refer to this as the ‘Cournot
problem’ since it was first articulated by Augustin Cournot ([1838] 1927, p. 127).
19
Some economists reserve the term ‘representative-agent models’ for models with a
single, infinitely-lived agent. In a typical overlapping-generations model the new young are
born at the start of every period, and the old die at the end of every period, and the model has
infinitely many periods; so there are infinitely many agents. On this view, the overlappinggenerations model is not a representative-agent model. I, however, regard it as one, because
within any period one type of young agent and one type of old agent stand in for the
enormous variety of people, and the same types are repeated period after period.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
FACTS AND ARTIFACTS
287
K.D.HOOVER
39
model permits better use of other information. Lars Peter Hansen and
Sargent (1980, pp. 91, 92), for example, argue that the strength of their
estimation method is that it accounts consistently for the interrelationships
between constituent parts of the model—i.e. that is a general-equilibrium
method. Calibrators respond, however, that it is precisely the importance of
general equilibrium that supports their approach. Kydland and Prescott write:
…it is in the stage of calibration where the power of the general equilibrium approach
shows up most forcefully. The insistence on internal consistency implies that
parsimoniously parameterized models of the household and business sector display
rich dynamic behavior through the intertemporal substitution arising from capital
accumulation and from other sources. (1991, p. 170)
The trade-off between the gains and losses of the two methods is not clear cut.
Lucas (1987, pp. 46, 47) and Prescott (1986, p. 15) argue that the strength of
calibration is that it uses multiple sources of information, supporting the belief that
it is structured around true invariants. This argument would appear to appeal to
the respectable, albeit philosophically controversial view, that a theory is better
supported when tested on information not used in its formulation (see Lipton
1991, ch. 8; Hoover 1994b). Unfortunately, it is not clear that calibration relies on
independent information nor that it avoids estimation altogether. Parameters are
sometimes chosen for calibrated business-cycle models because they mimic socalled ‘stylized facts’. That the models then faithfully reproduce such facts is not
independent information. Other parameters are chosen from microeconomic
studies. This introduces estimation through the back door, but without any but a
subjective, aesthetic metric to judge model performance.
Furthermore, since all new classical, equilibrium business-cycle models rely on
the idealization of the representative agent, both calibrated and estimated versions
share a common disability: using the representative-agent model in any form begs
the question by assuming that aggregation does not fundamentally alter the
structure of the aggregate model. Physics may provide a useful analogy. The laws
that relate pressure, temperature, and volumes of gases are macrophysics. The
‘ideal-gas laws’ can be derived from a micromodel: gas molecules are assumed to
be point masses, subject to conservation of momentum, with a distribution of
velocities. An aggregate assumption is also needed: the probability of the gas
molecules moving in any direction is taken to be equal.
Direct estimation of the ideal gas laws shows that they tend to break down—
and must be corrected with fudge factors—when pushed to extremes. For
example, under high pressures or low temperatures the ideal laws must be
corrected according to van der Waals’ equation. This phenomenal law, a law in
macrophysics, is used to justify alterations of the micromodel: when pressures
are high one must recognize that forces operate between individual molecules.
The inference of the van der Waals’ force from the macrophysical
behavior of gases has an analogue in the development of real-business-cycle
models. Gary Hansen (1985), for example, introduces the microeconomic
device of indivisible labor into a real-business cycle model, not from any direct
reflection on the nature of labor markets at the level of the individual firm or
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
288
40
METHODOLOGY
FACTS AND ARTIFACTS
worker, but as an attempt to account for the macroeconomic failure of
Kydland and Prescott’s (1982) model to satisfactorily reproduce the relative
variabilities of hours and productivity in the aggregate data.20 Of course,
direct estimation of Kydland and Prescott’s model rather than calibration may
have pointed in the same direction.21
Despite examples of macro to micro inference analogous to the gas laws,
Lucas’s (1980, p. 291) more typical view is that we must build our models up
from the microeconomic to the macroeconomic. Unlike gases, human society
does not comprise homogeneous molecules, but rational people, each choosing
constantly. To understand (verstehen) their behavior, one must model the individual
and his situation. This insight is clearly correct, it is not clear in the least that it is
adequately captured in the heroic aggregation assumptions of the representativeagent model. The analogue for physics would be to model the behavior of gases
at the macrophysical level, not as derived from the aggregation of molecules of
randomly distributed momenta, but as a single molecule scaled up to observable
volume—a thing corresponding to nothing ever known to nature.22
5. Calibration and macroeconomic practice
The calibration methodology has both a wide following and a substantial
opposition within the new classical school. I have attempted to give it a
sympathetic reading—both in general and in its specific application to realbusiness-cycle models. I have concentrated on Kydland and Prescott, as its
most prolific practitioners, and on Lucas, an articulate advocate. Although
calibration is consistent with appealing accounts of the nature and role of
models in science and economics, of their quanfication and idealization, its
practical implementation in the service of real-business-cycle models with
representative agents is less than compelling.
Does the calibration methodology amount to a repudiation of econometric
estimation altogether? Clearly not. At some level, econometrics still helps to
supply the values of the parameters of the models. Beyond that, whatever has
been said in favor of calibration methods to the contrary notwithstanding, the
20
Canova (1991b, p. 33) suggests that the particular covariance that Hansen’s modification of Kydland and Prescott’s model was meant to capture is an artifact of the HodrickPrescott filter, so that Hansen’s model may be a product of misdirected effort rather than a
progressive adaptation.
21
This, rather than collegiality, may account for Kydland and Prescott’s (1982, p. 1369)
tolerant remark about the future utility of Lars Peter Hansen and Sargent‘s (1980) econometric techniques as well as for Lucas’s (1987, p. 45) view that there is something to be learned
from Altug’s estimations of the Kydland and Prescott model—a view expressed in the midst of
arguing in favour of calibration.
22
A notable, non-new-classical attempt to derive macroeconomic behavior from
microeconomic behavior with appropriate aggregation assumptions is Durlauf (1989). In a
different, but related context, Stoker (1986) shows that demand systems fit the data only if
distributional variables are included in the estimating equations. He takes this macroeconomic evidence as evidence for the failure of the microeconomic conditions of exact aggregation.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
FACTS AND ARTIFACTS
K.D.HOOVER
289
41
misgivings of econometricians such as Sargent are genuine. The calibration
methodology, to date, lacks any discipline as stern as that imposed by
econometric methods. For Lucas (1980, p. 288) and Prescott (1983, p. 11), the
discipline of the calibration method comes from the paucity of free parameters.
But one should note that theory places only loose restrictions on the values of
key parameters. In practice, they are actually pinned down from econometric
estimation at the microeconomic level or accounting considerations. Thus, in
some sense, the calibration method would appear to be a kind of indirect
estimation. Thus, although as was pointed out earlier, it would be a mistake to
treat calibration as simply an alternative form of estimation, it is easy to
understand why some critics interpret it that way. Even were there less flexibility
in the parameterizations, the properties ascribed to the underlying components
of the idealized real-business-cycle models (the agents, their utility functions,
production functions, and constraints) are not subject to as convincing cross
checking as the analogous components in physical models usually are. The
fudge factors that account for the discrepancies between the ideal model and the
data look less like van der Waals’ equation than like special pleading. Above all,
it is not clear on what standards competing, but contradictory, models are to be
compared and adjudicated.23 Some such standards are essential if any objective
progress is to be made in economics.24
ACKNOWLEDGEMENTS
I thank Thomas Mayer, Kevin Salyer, Steven Sheffrin, Roy Epstein, Nancy
Cartwright, Gregor Smith, Edward Prescrott, Adrian Pagan, and two anonymous
referees for helpful comments on an earlier draft. The earliest version of this
paper, entitled ‘Calibration versus Estimation: Standards of Empirical Assessment
in the New Classical Macroeconomics’, was presented at the American Economic
Association meetings in Washington, DC, December 1990.
23
Prescott (1983, p. 12) seems oddly, to claim that inability of a model to account for
some real events is a positive virtue—in particular, that the inability of real-business-cycle
models to account for the Great Depression is a point in their favour. He writes: ‘If any
observation can be rationalized with some approach, then that approach is not scientific’.
This seems to be a confused rendition of the respectable Popperian notion that a theory is
more powerful the more things it rules out. But one must not mistake the power of a theory
with its truth. Aside from issues of tractability, a theory that rationalizes only and exactly
those events that actually occur, while ruling out exactly those events that do not occur is the
perfect theory. In contrast, Prescott seems inadvertently to support the view that the more
exceptions the better rule.
24
Watson (1993) develops a goodness-of-fit measure for calibrated models. It takes into
account that, since idealization implies differences between model and reality that may be systematic, the errors-in-variables and errors-in-equations statistical models are probably not appropriate. Also see Gregory and Smith (1991, pp. 27–8), Canova (1991a), and Hoover (1994a).
REFERENCES
ADELMAN, I. and ADELMAN, F.L. (1959). ‘The Dynamic Properties of the Klein-Goldberger
Model’, Econometrica, 27, 596–625.
ALDRICH, J. (1989). ‘Autonomy’, Oxford Economic Papers, 41, 15–34.
ALTUG, S. (1989). ‘Time-to-Build and Aggregate Fluctuations: Some New Evidence’,
International Economic Review, 30, 889–920.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
290
42
METHODOLOGY
FACTS AND ARTIFACTS
BAN SAL, R. and GALLANT, R.A., H U S S EY, R., and TAUCH E N, G. (1991).
‘Nonparametric Structural Estimation of Models for High-Frequency Currency Market
Data’, unpublished typescript.
BLAUG, M. (1992). The Methodology of Economics: Or How Economists Explain, Cambridge
University Press, Cambridge, 2nd ed.
CANOVA, F. (1991a). ‘Sensitivity Analysis and Model Evalution in Simulated Dynamic
General Equilibrium Economies’, unpublished typescript, Department of Economics,
European University Institute, Florence, Italy.
CANOVA, F. (1991b). ‘Detrending and Business Cycle Facts’, unpublished typescript,
Department of Economics, European University Institute, Florence, Italy.
CANOVA, F., FINN, M., and PAGAN, A.R. (1992). ‘Evaluating a Real Business Cycle
Model’, unpublished typescript.
COGLEY, T. and NASON, J.M. (1993). ‘Effects of the Hodrick-Prescott Filter on Trend and
Difference Stationary Time Series: Implications for Business Cycle Research’, unpublished
typescript.
COOLEY, T.F. and HANSEN, G.D. (1989). ‘Welfare Consequences of Monetary and Fiscal
Policy’, paper presented to the Federal Reserve Bank of San Francisco’s Fall Academic
Conference, 1 December.
COURNOT, A. [1838] (1927). Researches into the Mathematical Principles of the Theory of Wealth,
Nathaniel T.Bacon (trans,), Macmillan, New York.
DIAMOND, P.A. (1984). A Search-Equilibrium Approach to the Micro Foundations of Macroeconomics:
The Wicksell Lectures, 1982, MIT Press, Cambridge, MA.
DURLAUF, S.N. (1989). ‘Locally Interacting Systems, Coordination Failure, and the Behavior
of Aggregate Activity’ , unpublished typescript.
FRIEDMAN, M. (1953). ‘The Methodology of Positive Economics’, in Essays in Positive
Economics, Chicago University Press, Chicago, IL.
FRISCH, R. (1933). ‘Propagation Problems and Impulse Response Problems in Dynamic
Economics’, in Economic Essays in Honour of Gustav Cassel: October 20th 1933, George Allen
and Unwin, London.
GREGORY, A.W. and SMITH, G.W. (1990a). ‘Calibration as Estimation’, Econometric Reviews,
9, 57–89.
GREGORY, A.W. and SMITH, G.W. (1990b). ‘Calibration as Testing: Inference in Simulated
Macroeconomic Models’ in Journal of Business and Economic Statistics, forthcoming.
G REGORY, A.W. and S M ITH, G.W. (1991). ‘Calibration in Macroeconomics’, in
G.S.Maddala and C.R.Rao (eds), Handbook of Statistics, 10: Econometics, forthcoming.
HAAVELMO, T. (1944). ‘The Probability Approach in Econometrics’, Econometrica, 12
(supplement).
HALL, R.E. (1977). ‘Investment, Interest Rates, and the Effects of Stabilization Policies’
Brookings Papers on Economic Activity, 6, 61–100.
HANSEN, G.D. (1985). ‘Indivisible Labor and the Business Cycle’, Journal of Monetary
Economics, 16, 309–28.
HANSEN, G.D. and PRESCOTT, E.C. (1993). ‘Did Technology Shocks Cause the 1990–
1991 Recession’, American Economic Review, 83, 280–6.
HANSEN, G.D. and SARGENT, T.J. (1988), ‘Straight Time and Overtime in Equilibrium’,
Journal of Monetary Economics, 21, 281–308.
HANSEN, L.P. and SARGENT, T.J. (1980). ‘Formulating and Estimating Dynamic Linear
Rational Expectations Models’, in R.E.Lucas, Jr. and T.J.Sargent (eds), Rational Expectations
and Econometric Practice, George Allen & Unwin, London, 1981.
HARTLEY, J., SALYER, K., and SHEFFRIN, S. (1992). ‘Calibration and Real Business
Cycle Models: Two Unorthodox Tests’, unpublished typescript.
HOOVER, K.D. (1984). ‘Two Types of Monetarism’, Journal of Economic Literature, 22, 58–76.
HOOVER, K.D. (1988). The New Classical Macroeconomics: A Skeptical Inquiry, Blackwell,
Oxford.
HOOVER, K.D. (1992). ‘The Rational Expectations Revolution: An Assessment’, The Cato
Journal, 12, 81–96.
HOOVER, K.D. (1994a). ‘Six Queries About Idealization in an Empirical Context’, Poznan
Studies in the Philosphy of Science and the Humanities, 38, 43–53.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
FACTS AND ARTIFACTS
K.D.HOOVER
291
43
HOOVER, K.D. (1994b). ‘In Defence of Datamining: Some Preliminary Throughts’, in
K.D.Hoover and S.M.Sheffrin (eds), Monetarism and the Methodology of Econmics: Essays in
Honor of Thomas Mayer, Elgar, Aldershot, forthcoming.
KING, R.G. and PLOSSER, C.I. (1984). ‘Money, Credit and Prices in a Real Business
Cycle’, American Economic Review, 74, 363–80.
KING, R.G. and PLOSSER, C.I. (1989). ‘Real Business Cycle Models and the Test of the
Adelmans’, paper presented to the Federal Reserve Bank of San Francisco’s Fall Academic
Conference. 1 December 1989.
KOOPMANS, T.C. (1947). ‘Measurement without Theory’, Review of Economics and Statistics,
29, 161–72.
KYDLAND, F.E. (1984). ‘Labor-force Heterogeneity and the Business Cycle’ in K.Brunner
and A.H.Meltzer (eds), Essays on Macroeconomic Implications of Financial and Labor Markets
and Political Processes, Carnegie-Rochester Conference Series on Public Policy, 21.
KYDLAND, F.E. and PRESCOTT, E.C. (1982). ‘Time to Build and Aggregate Fluctuations’,
Econometrica, 50, 1345–70.
KYDLAND, F.E. and PRESCOTT, E.C. (1988). ‘The Workweek of Capital and Its Cyclical
Implication’ Journal of Monetary Economics, 21, 243–60.
KYDLAND, F.E. and PRESCOTT, E.C. (1990). ‘Business Cycles: Real Facts and a Monetary
Myth’, Federal Reserve Bank of Minneapolis Quartly Review, 14, 3–18.
KYDLAND, F.E. and PRESCOTT, E.C. (1991). ‘The Econometrics of the Great Equlibrium
Approach to Business Cycles’, Scandinavian Journal of Economics, 93, 161–78.
LAKATOS, I. (1970). ‘Falsification and the Methodology of Scientific Research Programmes’,
in I. Lakatos and A.Musgrave (eds), Criticism and the Growth of Knowledge, Cambridge
University Press, Cambridge.
LAKATOS, L (1978). ‘History of Science and Its Rational Reconstructions’ in J.Worrall and
G. Currie (eds), The Methodology of Scientific Research Programmes: Philosphical Papers, I,
Cambridge University Press, Cambridge.
LAU, L.J. (1984). ‘Comments’, in H.E.Scarf and J.B.Shoven (eds), Applied General Equilibrium
Analysis, Cambridge University Press, Cambridge.
LIPTON. P. (1991). Inference to the Best Explanation, Routlege, London.
LUCAS, R.E., JR. (1976). ‘Econometric Policy Evaluaton: A Critique’, in Studies in BusinessCycle Theory, Oxford, Blackwell, 1981.
LUCAS, R.E., JR. (1977). ‘Understanding Business Cycles’, in Studies in Business-Cycle Theory,
Blackwell, Oxford, 1981.
LUCAS, R.E.. JR. (1980) ‘Methods and Problems in Business Cycle Theory’, in Studies in
Business-Cycle Theory, Blackwell, Oxford, 1981.
LUCAS, R.E. JR. (1986). ‘Adapttive Behavior and Economic Theory’, Journal of Business,
59, S401–26.
LUCAS. R.E., JR. (1987), Models of Business Cycles, Blackwell, Oxford.
LUCAS, R.E., JR. and SARGENT, T.J.C. (eds) (1981). Rational Expectations and Econometric
Practice, George Allen and Unwin, London.
MACKINNON.J. (1984). ‘Comments’, in H.E.Scarf and J.B.Shoven (eds), Applied General
Equilibrium Analysis , Cambridge University Press, Cambridge.
MAN SU R, A. and WALLEY, J. (1984). ‘Numerical Specification of Applied Generalequilibrium models: Estimation, Calibration, and Data’, in H.E.Scarf and J.B.Shoven
(eds.), Applied General Equilibrium Analysis, Cambridge University Press, Cambridge.
MANUELLI, R. and SARGENT.T.J. (1988). ‘Models of Business Cycles: A Review Essay’
Journal of Monetary Economics, 22, 523–42.
MAYER, T. (1960). ‘Plant and Equipment Lead Times’ Journal of Business, 33, 127–32.
MCCALLUM, B.T. (1986). ‘On “Real” and “Sticky-Price” Theories of the Business Cycle’,
Journal of Money, Credit and Banking, 18, 397–414.
MCCALLUM, B.T. (1989). ‘Real Business Cycle Models’, in R.J.Barro (ed.), Modern Business
Cycle Theory, Blackwell, Oxford.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
Part V
The critique of calibration
methods
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
CHAPTER 16
© 1991 American Statistical Association
295
Journal of Business & Economic Statistics, July 1991, Vol. 9, No. 3
Calibration as Testing: Inference in
Simulated Macroeconomic Models
Allan W.Gregory and Gregor W.Smith
Department of Economics, Queen’s University, Kingston, Ontario, K7L 3N6, Canada
A stochastic macroeconomic model with no free parameters can be tested by comparing its features, such as
moments, with those of data. Repeated simulation allows exact tests and gives the distribution of the sample
moment under the null hypothesis that the model is true. We calculate the size of tests of the model studied
by Mehra and Prescott. The approximate size of their test (which seeks to match model-generated, mean, riskfree interest rates and equity premia with historical values) is 0 although alternate, empirical representations
of this model economy or alternate moment-matching tests yield large probabilities of Type I error.
KEY WORDS: Equity premium; Monte Carlo; Simulation; Type I error.
Calibration in macroeconomics is concerned
primarily with testing a model by comparing
population moments (or perhaps some other
population measure) to historical sample moments of actual data. If the correspondence between some aspect of the model and the historical record is deemed to be reasonably close,
then the model is viewed as satisfactory. If the
distance between population and historical moments is viewed as too great, then the model is
rejected, as in the widely cited equity-premium
puzzle of Mehra and Prescott (1985). A drawback to the procedure as implemented in the
literature is that no metric is supplied by which
closeness can be judged. This leads to tests with
unknown acceptance and rejection regions.
This article provides a simple way to judge the
degree of correspondence between the population
moments of a simulated macroeconomic model
and observed sample moments and develops a
framework for readily calculating the size
(probability of Type I error) of calibration tests. We
apply this method to the well-known equitypremium case. This article is not concerned with a
“solution” to the equity-premium puzzle. Rather it
evaluates the probability of falsely rejecting a true
macroeconomic model with calibration methods.
One finding is that the size of the test considered by
Mehra and Prescott (which seeks to match mean
risk-free interest rates and equity premia) is 0, so the
model with their parameter settings is unlikely to
have generated the observed historical moments.
Some alternate versions of the consumption-based
asset-pricing model or alternate moment-matching
tests yield large probabilities of Type I error.
Section 1 characterizes calibration as testing.
A simple formalization of calibration as Monte
Carlo testing allows exact inference. Section 2
contains an application to the test conducted by
Mehra and Prescott (1985). Section 3 concludes.
1. CALIBRATION AS TESTING
Calibration in macroeconomics has focused on
comparing observed historical moments with
population moments from a fully parameterized
simulation model—that is, one with no free parameters. One might elect to simulate a model
because of an analytical intractability or because a forcing variable is unobservable. In
macroeconomics, examples of unobservable
forcing variables include productivity shocks in
business-cycle models or consumption measurement errors in asset-pricing models.
Consider a population moment θ, which is
restricted by theory, with corresponding
historical sample moment for a sample of size
T. Call the moment estimator . Assume that
is consistent for θ. The population moment is a
number, the sample moment is the realization of
a random variable (an estimate), and the
estimator is a random variable. The calibration
tests applied in the literature compare θ and
and reject the model if θ is not sufficiently close
to . In some calibration studies, attempts are
made to exactly match the population moment
to the sample moment (there must be some
leeway in parameter choice to make this
attempt nontrivial). Such matching imposes
unusual test requirements because θ and can
differ even when the model is true due to
sampling variability in . Moreover, judging
closeness involves the sampling distribution of
the estimator . Standard hypothesis testing
procedures may be unavailable because the
exact or even asymptotic distribution of the
estimator is unknown.
One prominent advantage in the calibration
of macroeconomic models that has not been
exploited fully is that the complete datagenerating process is specified. Thus the
297
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
296
298
CELEBRATION
Journal of Business & Economic Statistics, July 1991
sampling variability of the simulated moment can
be used to evaluate the distance between θ and
Call N the number of simulated observations.
We construct tests by repeatedly simulating the
fully parameterized model (or artificial economy)
and calculating the proportion of times lies in
a set ⍜(bounded by ). Current calibration
studies do simulate repeatedly, but they typically
average over a small number of replications and
then quote the averages of various properties.
Our approach involves treating the historical
as a critical value to give the
moment
probability (prob) value (i.e., marginal
significance level) or size of the test implicit in
comparing moments. If simulations use the same
number of observations as are used in calculating
the historical sample moment so that N=T, then
the test will be exact. Moreover, confidence
intervals can be constructed for (as opposed to
the population moment θ ) to determine whether
lies within the interval.
Barnard (1963) suggested basing exact
inference on repeated generation of artificial
series. Monte Carlo testing was refined by Hope
(1968) and Marriott (1979) and applied in
economics by Theil, Shonkwiler, and Taylor
(1985) and Theil and Shonkwiler (1986).
Simulation methods also can be used for
estimation, as described for macroeconomic
models by Lee and Ingram (1991) and Gregory
and Smith (1990).
Suppose one wishes to find the probability
with which a model with population moment θ
generates a sample moment
Our
suggested procedure is to simulate N
observations from the economy, with N=T, R
times and calculate at each replication. The
proportion of replications with which the
simulated moment exceeds the corresponding
sample moment is the (one-sided) prob value
associated with the historical moment
(1)
in which the indicator function I is defined to be
equal to 1 if its argument is positive and 0 otherwise.
Simulations involve a finite number of
replications, so measures of test size or
confidence intervals are themselves random
variables. Standard errors for test size may be
calculated from the binomial distribution or, if
R is large, the normal distribution. Although we
have described treating the historical moments
as critical values, one also could select a test size
and estimate the corresponding critical value
(i.e., the quantile) from the order statistics of the
Monte Carlo sample of moments.
In some cases one might be interested in the
entire probability density of the sample moment
under the null hypothesis that the model is true.
Although it is simpler to work with the empirical
density function, tests also could be based on the
density estimated nonparametrically from
], and approximate
simulations [denoted by
measures of size could be calculated. There are
many estimators available (see Silverman 1986
or Tapia and Thompson 1978), but if we assume
that the density is absolutely continuous, then
kernel estimators provide an easy way to smooth
the empirical density. One advantage in applying
kernel estimators is that they are available in an
IMSL subroutine.
In the application in Section 2, we calculate
prob values in two ways. The first way is the
simplest and uses the empirical distribution as in
Equation (1). The second way estimates the
density nonparametrically with a quartic kernel
and, following Silverman (1986), a robust,
variable window width given by .698 min [s,
IQR/ 1.34], in which s is the empirical standard
deviation and IQR the interquartile range. Under
weak conditions the empirical distribution
function and the nonparametric density
consistently estimate the true cumulative
distribution function and density, respectively. In
the simulations in Section 2, the results from the
two methods usually are identical, which suggests
using the simpler calculation in Equation (1).
As N becomes large, sample moments
converge to population moments by ergodicity.
For fixed N, sample moments have distributions;
as the number of replications R increases, the
estimated distribution function converges to the
true finite-sample distribution. One could
examine the sensitivity of findings to alternative
windows, kernels, and numbers of replications
because the speed of convergence may be slow.
For example, many replications may be
necessary for accurate estimates of tail-area
probabilities (i.e., test size), where by definition
events are rare in a sampling experiment.
Finally, one also could study tests based on
matching properties other than moments, by
analogy with the estimation results of Smith
(1989), who considered matching properties of
misspecified models or linear regressions in
simulated and historical data. Naylor and Finger
(1971, p. 159) listed other validation criteria,
such as matching properties of turning points.
2. APPLICATION: TYPE I ERROR IN
THE EQUITY-PREMIUM PUZZLE
As an example, consider the exchange economy
described by Mehra and Prescott (1985). A
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
CALIBRATION AS TESTING
297
Gregory and Smith: Calibration as Testing 299
representative consumer has preferences characterized by the utility functional
(2)
in which Et denotes the expectation conditional
on information at time t, β 僆 (0,1) is a discount
factor, u(c) =c1-α/(1-α) [log(c) if α=1], and α is a
positive, constant coefficient of relative risk
aversion. The consumer’s nonstorable endowment yt evolves exogenously according to
(5b)
Thus if the current state is (yt, li), then the prices
(relative to one unit of the commodity at time t)
of the two assets are
(6a)
and
(3)
(6b)
in which the growth rate, xt, follows a process that is
Markov on a discrete state space Λ={λ1, λ2, …, λJ}.
This process is stationary and ergodic, with
transition matrix ␾ ,
(4)
The equilibrium or unconditional probabilities
are given by φi=Pr[xt=λi] for all t.
An equilibrium in this economy is
characterized by a set of prices for which
consumption equals the endowment at all times
and in all states. Relative prices are calculated
by equating them to marginal rates of
substitution. The price of a one-period, risk-free
discount bond that provides one unit of the
endowment at time t+1 is given by
(5a)
and the price of an equity claim to the endowment stream is given by
The challenge posed of this model by Mehra and
Prescott has been the following: With a risk-aversion parameter α between 0 and 10, a discount
factor β between 0 and 1, and the Markov per
capita consumption growth-rate process matching the sample mean, variance, and first-order
autocorrelation of the U.S. series for 1889–
1978, can the model (with two or three states)
generate a population mean risk-free rate of return and mean equity premium that match the
annual U.S. sample values (.8% and 6.2%, respectively) of these measures? Mehra and
Prescott compared the 6.2% average equity premium from the full sample (90 observations) to
an average premium from their model of at most
.35%. These two values were judged not to be
close, and hence there is a puzzle.
One could try to gauge closeness with a
standard asymptotic hypothesis test. For example,
the test statistic implicit in Mehra and Prescott’s
(1985, p. 156) quotation of the estimated average
equity premium and its standard error is
Table 1. Asset-Pricing Models and Joint Tests
NOTE: The prob value is the marginal significance level of the joint one-sided test that the population
mean equity premium is at least 6.2% and the population mean risk-free rate is less than 4%.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
298
300
CELEBRATION
Journal of Business & Economic Statistics, July 1991
Table 2. Sample and Population Moments: Case 1, Symmetric (Mehra-Prescott)
NOTE: The symbols µ. σ, and ρ denote mean, variance, and first-order autocorrelation. Sample
moments in the third column are based on the full sample. The sample autocorrelations for the riskfree and equity-premium rates were not given in Mehra and Prescottís table 1 (1985, p. 147). Pop.
denotes population moments from the model. Emp denotes measures constructed from the empirical
density function and Np those constructed from the nonparametrically estimated density function,
each based on R=1,000 replications. Confidence intervals run from the .025 to als run from the .025
to .975 quantiles.
a Prob values for
b Prob values for
asymptotically. In
this test one could correct the standard error of
the average equity premium to allow for
dependence and heterogeneity with the methods
of Newey and West (1987) or Andrews (1988).
The heterogeneity does not matter in a test with
no regressors (as in a comparison of means), but
the serial correlation does.
Cecchetti, Lam, and Mark (1990) constructed
an asymptotic test statistic for matching a moment:
where is
W=
the Newey-West estimate of the variance of the
sample moment. To allow for uncertainty about
parameters that are estimated, they assumed that
the two sources of uncertainty are independent
and amended the statistic to W=
where
is the variance due to parameter
uncertainty. Kocherlakota (1990) conducted a
Monte Carlo experiment of tests of
consumption-based asset-pricing models that
use asymptotic 5% critical values and examined
the small-sample size of such tests with simulated
data. He found that size most closely
approximates asymptotic size when parameters
are set rather than estimated. Finding a
consistent estimate even of the asymptotic
standard error for many estimated moments can
be very difficult, however. Moreover, some of
the historical samples considered by Mehra and
Prescott have as few as 10 observations and
hence should not be studied with asymptotic
methods. On the other hand, the procedure
outlined in Section 1 is appropriate for inference
in such sample sizes using the sampling
Table 3. Sample and Population Moments: Case 2, Crash (Rietz)
NOTE: See note to Table 2.
a
Prob values for
b
Prob values for
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
CALIBRATION AS TESTING
299
Gregory and Smith: Calibration as Testing 301
Table 4. Sample and Population Moments: Case 3, Asymmetric
NOTE. See note to Table 2.
a
Prob values for
b
Prob values for
variability inherent in fully parameterized
models.
To investigate the size of the momentmatching exercise considered by Mehra and
Prescott, suppose that their model is true. What is
the probability of observing an equity premium
of at least 6.2%? To answer this question, we
simulate the fully parameterized model, estimate
the probability density functions for the random
variables that it restricts, and measure tail-area
probabilities. We follow the same procedure for
four different representations of the transition
density for the Markov chain and of risk
aversion. These are shown in Table 1. The first
representation is that considered by Mehra and
Prescott. The second representation is that used
by Rietz (1988), in which there is a small
probability of a “crash” in growth rates; the
source is Rietz’s example 1, table 3, row 1. The
third representation involves an asymmetry in
growth transitions, and the fourth representation
involves a degree of risk aversion greater than
that considered by Mehra and Prescott.
The explicit joint test considered by Mehra
and Prescott (1985, p. 154) is to generate a
population equity premium of at least 6.2% and a
population risk-free rate less than 4%. For each
model we estimate the prob value associated
with this joint test. These measures are based on
bivariate empirical distributions with R= 1,000
replications and on N=90 observations, just as in
the annual sample. The results are shown in
Table 1. For the Mehra-Prescott model of case 1,
the prob value is 0. The other models generate
the historical sample-mean rates of return with
positive probabilities. The prob values are .59
for case 2, .14 for case 3, and .95 for case 4.
Tables 2–5 give the sample mean, variance,
and first-order autocorrelation of consumption
growth, the risk-free rate, and the equity
premium for the annual data used by Mehra
and Prescott and the corresponding population
values for the four simulation models described
in Table 1. The autocorrelations of the risk-free
rate and consumption growth are equal as an
analytical property of the model. For each
Table 5. Sample and Population Moments: Case 4, Extreme Risk Aversion
NOTE: See note to Table 2.
a
Prob values for
b
Prob values for
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
300
302
CELEBRATION
Journal of Business & Economic Statistics, July 1991
Figure 1. The Nonparametrically Estimated Density
Function (based on 1,000 replications) of the Mean
Equity Premium in Simulated Samples of 90 Observations, From the Model Economy (4), With Extreme Risk
Aversion. Parameter settings are given in Table 1.
model we construct 95% confidence intervals
for a sample size of 90 observations. We also
estimate the densities nonparametrically and
calculate corresponding approximate intervals.
Tables 2–5 also give size results for some
univariate, one-sided tests of each model’s
ability to match individual moments.
For example, Table 2 shows that for one of the
Mehra-Prescott parameter settings (given in Table
1) the equity premium has a mean of .20. The
estimated probability of observing a premium
greater than 6.2% is 0; the 95% confidence
interval is (-.6, 1.0) at N=90 observations. One
can also see that the puzzle extends to second
moments, because the variances of historical riskfree rates and equity premia lie outside the
confidence intervals generated by the theoretical
model. Mehra and Prescott (1985, p. 146) did
note that their model was poorly suited for assetprice volatility issues. The confidence intervals
for autocorrelation coefficients are generally wide
and would rationalize a wide range of observed
sample values.
The Rietz model (2) in Table 3 is consistent
with the sample mean equity premium of 6.18%
and, in fact, would be consistent with an even
larger premium; the 95% confidence interval is
(5.1, 7.5) at N=90 observations and the prob value
(from the simulated empirical density) is .59. Like
model 1, it fails to match the variances of the riskfree rate and equity premium. The model with
asymmetric growth (3) in Table 4 generates larger
variances for the two returns but still cannot match
the variance of the equity-premium rate. Model 3
is consistent with the sample mean-equity
premium in that it generates the sample value with
a prob value of .10, but it also gives a consumption
growth variance that is too large relative to the
sample variance. In Table 5, model 4, in which
α=25, gives 95% confidence intervals that include
the historical values for all moments except the
variance of the risk-free rate. Although models 1
and 2 underpredict the variance of the risk-free
rate, models 3 and 4 overpredict it. Finally, for
model 4, Figure 1 illustrates the nonparametrically
estimated density function of the mean equity
premium from which the confidence interval in
the last column of Table 5 is taken.
In Tables 2–5 the confidence intervals and size
estimates from the empirical densities and the
nonparametrically estimated densities are virtually
identical. We also have used the same method with
a sample size of N=10 for comparison with sample
values reported by Mehra and Prescott (1985,
table 1) for 1959–1968. Naturally the intervals for
the smaller simulated sample size are wider. We
also have investigated the effects of increasing the
number of replications to R=2,000. This increase
has no appreciable effect on the confidence
intervals. Although faster convergence in R might
be possible with the use of control or antithetic
variates to reduce sampling variability in estimated
test sizes and confidence intervals, such methods
do not seem vital in this application.
In constructing the tests in the tables, we do not
rule out applying other criteria to evaluate versions
of this asset-pricing model. For example, one might
discard the fourth case because of its large risk
aversion or the second case because its growthtransition matrix includes elements never
observed (see Mehra and Prescott 1988). We are
agnostic on these points. The proposed procedure
simply allows one to gauge the sampling
variability of the moment estimator when the
model is assumed to be true and thus permits
formal comparisons of historical and population
(model) moments. Kwan (1990) and Watson
(1990) have proposed alternative measures of the
goodness of fit of calibrated models (or closeness of
moments) in which the parameterized economic
model is not the null hypothesis.
3. CONCLUSION
Calibration in macroeconomics involves simulating dynamic models and comparing them to
historical data. This type of comparison can be
seen as a test of the model. A natural issue in the
use of this method concerns the distribution of the
sample moment under the null hypothesis that
the model is true. For fully parameterized model
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
CALIBRATION AS TESTING
301
Gregory and Smith: Calibration as Testing 299
economies, one can use repeated simulation to
estimate the probability of falsely rejecting a true
economic model (Type I error) and to construct
confidence intervals using comparisons of moments. The empirical frequency estimator works
well in both cases and is very simple to apply.
As an application we consider the assetpricing model and parameter settings of Mehra
and Prescott, for which the test of mean rates of
return has no probability of Type I error; their
rejection of the model appears to be statistically
sound. Our results can be used to find the sizes of
other moment-matching tests of this model. If
other parameter settings are admitted, then the
model will generate an equity premium of 6.2%
more than 5% of the time. The method can be
applied directly to extensions of the asset-pricing
economy discussed previously (see, for example,
Abel 1990; Benninga and Protopapadakis 1990;
Cecchetti et al. 1990; Labadie 1989) and to
business-cycle models, as Devereux, Gregory,
and Smith (1990) demonstrated.
ACKNOWLEDGMENTS
We thank the Social Sciences and Humanities
Research Council of Canada for financial support. We thank David Backus, Jean-Marie
Dufour, James MacKinnon, Adrian Pagan,
Simon Power, Tom Rietz, Jeremy Rudin, the editor, an associate editor, a referee, and numerous
seminar participants for helpful comments.
[Received April 1990. Revised January 1991.]
REFERENCES
Abel, A.B. (1990), “Asset Prices Under Habit
Formation and Catching up With the Joneses,”
American Economic Review, 80, 38–42.
Andrews, D.W.K. (1988), “Heteroskedasticity and
Autocorrelation Consistent Covariance Matrix
Estimation,” Cowles Foundation Discussion Paper
877R, Yale University, Dept. of Economics.
Barnard, G.A. (1963), Discussion of “The Spectral
Analysis of Point Process,” by M.S.Bartlett, Journal
of the Royal Statistical Society, Ser. B, 25, 294.
Benninga, S., and Protopapadakis, A. (1990),
“Leverage, Time Preference, and the ‘Equity
Premium Puzzle,’” Journal of Monetary Economics,
25, 49–58.
Cecchetti, S.G., Lam, P., and Mark, N.C. (1990),
“The Equity Premium and the Risk Free Rate:
Matching the Moments,” mimeo, Ohio State
University, Dept. of Economics.
Devereux, M.B., Gregory, A.W., and Smith, G.W.
(1990), “Realistic Cross-Country Consumption
Correlations in a Two-County Equilibrium
Business Cycle Model,” IER Discussion Paper
774, Queen’s University, Dept. of Economics.
Gregory, A.W., and Smith, G.W. (1990),
“Calibration as Estimation,” Econometric Reviews, 9,
57–89.
Hope, A.C.A. (1968), “A Simplified Monte Carlo
Significance Test Procedure,” Journal of the Royal
Statistical Society, Ser. B, 30, 582–598.
Kocherlakota, N.R. (1990), “On Tests of
Representative Consumer Asset Pricing Models,”
Journal of Monetary Economics, 26, 285–304.
Kwan, Y.-K. (1990), “Bayesian Model Calibration
With An Application to a Non-Linear Rational
Expectations Two-Country Model,” mimeo.
University of Chicago, Graduate School of Business.
Labadie, P. (1989), “Stochastic Inflation and the
Equity Premium,” Journal of Monetary Economics,
24, 277–298.
Lee, B.-S., and Ingram, B.F. (1991), “Simulation
Estimation of Time Series Models,” Journal of
Econometrics, 47, 197–205.
Marriott, F.H.C. (1979), “Barnard’s Monte Carlo
Tests: How Many Simulations?” Applied Statistics,
28, 75–77.
Mehra, R., and Prescott, E.C. (1985), “The Equity
Premium: A Puzzle,” Journal of Monetary Economics,
15, 145–161.
——(1988), “The Equity Risk Premium: A Solution?”
Journal of Monetary Economics, 22, 133–136.
Naylor, T.H., and Finger, J.M. (1971), “Validation,”
in Computer Simulation Experiments With Models of
Economic Systems, ed. T.H.Naylor, New York: John
Wiley, pp. 153–164.
Newey, W.K., and West, K.D. (1987), “A Simple,
Positive Semidefinite, Heteroskedasticity and
Autocorrelation Consistent Covariance Matrix,”
Econometrica, 55, 703–708.
Rietz, T. (1988), “The Equity Risk Premium: A
Solution,” Journal of Monetary Economics, 22, 117–
131.
Silverman, B.W. (1986), Density Estimation for Statistics
and Data Analysis, London: Chapman & Hall.
Smith, A. (1989), “Estimation of Dynamic
Economic Models by Simulation: A Comparison
of Two Approaches,” mimeo. Duke University,
Dept. of Economics.
Tapia, R.A., and Thompson, J.R. (1978),
Nonparameiric Probability Density Estimation,
Baltimore: The Johns Hopkins University Press.
Theil, H., and Shonkwiler, J.S., (1986), “Monte
Carlo Tests of Autocorrelation,” Economics Letters,
20, 157–160.
Theil, H., Shonkwiler, J.S., and Taylor, T.G. (1985).
“A Monte Carlo Test of Slutsky Symmetry,”
Economics Letters, 19, 331–332.
Watson, M.W. (1990), “Measures of Fit for
Calibrated Models,” mimeo, Northwestern
University, Dept. of Economics.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
302
CHAPTER 17
Measures of Fit for Calibrated Models
Mark W.Watson
Northwestern University and Federal Reserve Bank of Chicago
This paper suggests a new procedure for evaluating the fit of a dynamic structural
economic model. The procedure begins by augmenting the variables in the
model with just enough stochastic error so that the model can exactly match
the second moments of the actual data. Measures of fit for the model can then
be constructed on the basis of the size of this error. The procedure is applied to
a standard real business cycle model. Over the business cycle frequencies, the
model must be augmented with a substantial error to match data for the
postwar U.S. economy. Lower bounds on the variance of the error range from
40 percent to 60 percent of the variance in the actual data.
I. Introduction
Economists have long debated appropriate methods for assessing the
empirical relevance of economic models. The standard econometric
approach can be traced back to Haavelmo (1944), who argued that an
economic model should be embedded within a complete probability
model and analyzed using statistical methods designed for conducting
inference about unknown probability distributions. In the modern
literature, this approach is clearly exemplified in work such as that of
L.Hansen and Sargent (1980) or McFadden (1981). However, many
economic models do not provide a realistic and complete
This paper has benefited from constructive comments by many seminar participants; in
particular I thank John Cochrane, Marty Eichenbaum, Jon Faust, Lars Hansen, Robert
Hodrick, Robert King, and Robert Lucas. Two referees also provided valuable constructive
criticism and suggestions. The first draft of this paper was written while I was visiting the
University of Chicago, whose hospitality is gratefully acknowledged. This research was
supported by the National Science Foundation through grants SES-89–10601 and SES-9122463.
[Journal of Political Economy, 1993, vol. 101, no. 6]
© 1993 by The University of Chicago. All rights reserved. 0022–3808/93/0106–0007$01.50
1011
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
MEASURES OF FIT
1012
303
JOURNAL OF POLITICAL ECONOMY
probability structure for the variables under consideration. To analyze
these models using standard econometric methods, they must first be
augmented with additional random components. Inferences drawn from
these expanded models are meaningful only to the extent that the additional
random components do not mask or change the salient features of the
original economic models.
Another approach, markedly different from the standard econometric
approach, has become increasingly popular for evaluating dynamic
macroeconomic models. This approach is clearly articulated in the work
of Kydland and Prescott (1982) and Prescott (1986). In a general sense,
the approach asks whether data from a real economy share certain
characteristics with data generated by the artificial economy described by
an economic model. There is no claim that the model explains all the
characteristics of the actual data, nor is there any attempt to augment the
model with additional random components to more accurately describe
the data. On the one hand, the results from this approach are easier to
interpret than the results from the standard econometric approach since
the economic model is not complicated by additional random elements
added solely for statistical convenience. On the other hand, since the
economic model does not provide a complete probability structure, inference
procedures lack statistical foundations and are necessarily ad hoc. For
example, a researcher may determine that a model fits the data well because
it implies moments for the variables under study that are “close” to the
moments of the actual data, even though the metric used to determine the
distance between the moments is left unspecified.
This paper is an attempt to put the latter approach on a less ad hoc
foundation by developing goodness-of-fit measures for the class of dynamic
econometric models whose endogenous variables follow covariance
stationary processes. It is not assumed that the model accurately describes
data from the actual economy; the economic model is not a null hypothesis
in the statistical sense. Rather, the economic model is viewed as an
approximation to the stochastic processes generating the actual data, and
goodness-of-fit measures are proposed to measure the quality of this
approximation. A standard device—stochastic error—is used to motivate
the goodness-of-fit measures. These measures answer the question, How
much random error would have to be added to the data generated by the
model so that the autocovariances implied by the model+error match the
autocovariances of the observed data?
The error represents the degree of abstraction of the model from the
data. Since the error cannot be attributed to a data collection procedure or
to a forecasting procedure, for instance, it is difficult a priori to say much
about its properties. In particular, its covariance with the observed data
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
304
CELEBRATION
MEASURES OF FIT
1013
cannot be specified by a priori reasoning. Rather than make a specific
assumption about the error’s covariance properties, I construct a
representation that minimizes the contribution of the error in the complete
model. Thus, in this sense, the error process is chosen to make the model
as close to the data as possible.
Many of the ideas in this paper are close to, and were motivated by,
ideas in Altug (1989) and Sargent (1989). Altug (1989) showed how a
one-shock real business cycle model could be analyzed using standard
dynamic econometric methods, after first augmenting each variable in
the model with an idiosyncratic error. This produces a restricted version
of the dynamic factor analysis or unobserved index models developed by
Sargent and Sims (1977) and Geweke (1977). Sargent (1989) discusses two
models of measurement error: in the first the measurement error is
uncorrelated with the data generated by the model, and in the second the
measurement error is uncorrelated with the sample data (see also G.Hansen
and Sargent 1988). While similar in spirit, the approach taken in this
paper differs from that of Altug and Sargent in two important ways. First,
in this paper, the error process is not assumed to be uncorrelated with the
model’s artificial data or with the actual data. Rather, the correlation
properties of the error process are determined by the requirement that the
variance of the error is as small as possible. Second, the joint data-error
process is introduced to motivate goodness-of-fit measures; it is not
introduced to describe a statistical model that can be used to test statistical
hypotheses, at least in the standard sense. Rather, the analysis in this
paper is similar to the analysis in Campbell and Shiller (1988), Durlauf
and Hall (1989), Hansen and Jagannathan (1991), and Cochrane (1992).
Each of these papers uses a different approach to judge the goodness of fit
of an economic model by calculating a value or an upper bound on the
variance of an unobserved “noise” or a “marginal rate of substitution” or
a “discount factor” in observed data.
The minimum approximation error representation developed in this
paper motivates two sets of statistics that can be used to evaluate the
goodness of fit of the economic model. First, like the variance of the error
in a regression model, the variance of the approximation error can be
used to form an R2 measure for each variable in the model. This provides
an overall measure of fit. Moreover, spectral methods can be used to calculate
this R2 measure for each frequency so that the fit can be calculated over
the “business cycle,” “growth,” or other specific frequency bands. A second
set of statistics can be constructed by using the minimum error
representation to form fitted values of the variables in the economic model.
These fitted values show how well the model explains specific historical
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
MEASURES OF FIT
1014
305
JOURNAL OF POLITICAL ECONOMY
episodes; for example, can a real business cycle model simultaneously
explain the growth in the United States during the 1960s and the 1981–
82 recession?
The paper is organized as follows. Section II develops the minimum
approximation error representation and goodness-of-fit measures. Section
III calculates these goodness-of-fit statistics for a standard real business
cycle model using postwar U.S. macroeconomic data on output,
consumption, investment, and employment. Section IV concludes the
paper by providing a brief discussion of some tangential issues that arise
from the analysis.
II. Measures of Fit
Consider an economic model that describes the evolution of an n×1 vector
of variables xt. Assume that the variables have been transformed, say by
first-differencing or forming ratios, so that xt is covariance stationary. As
a notational device, it is useful to introduce the au toco variance generating
function (ACGF) of xt, Ax(z). This function completely summarizes the
unconditional second-moment properties of the process. In what follows,
“economic model” and “Ax(z)” will be used interchangeably; that is, the
analysis considers only the unconditional second-moment implications
of the model. Nonlinearities and variation in conditional second and
higher moments are ignored to help keep the problem tractable. The
analysis will also ignore the unconditional first moments of xt; modifying
the measures of fit for differences in the means of the variables is
straightforward.
The empirical counterparts of xt are denoted yt . These variables differ
from xt in an important way. The variables making up xt correspond to the
variables appearing in the theorist’s simplification of reality; in a
macroeconomic model they are variables such as “output,” “money,” and
the “interest rate.” The variables making up yt are functions of raw data
collected in a real economy; they are variables such as “per capita gross
domestic product in the United States in 1987 dollars” or “U.S. M2” or
“the yield on 3-month U.S. Treasury bills.”
The question of interest is whether the model generates data with
characteristics similar to those of the data generated by the real economy.
Below, goodness-of-fit measures are proposed to help answer this question.
Before I introduce these new measures, it is useful to review standard
statistical goodness-of-fit measures to highlight their deficiencies for
answering the question at hand.
Standard statistical goodness-of-fit measures use the size of sampling
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
306
CELEBRATION
MEASURES OF FIT
1015
error to judge the coherence of the model with the data. They are based
on the following: First, Ay(z), the population ACGF for yt, is unknown
but can be estimated from sample data. Discrepancies between the estimator
and arise solely from sampling error in
and the likely size of
the error can be deduced from the stochastic process that generated the
sample. Now, if Ay(z)=Ax(z), sampling error also accounts for the differences
and Ax(z). Standard goodness-of-fit measures show how
between
likely it is that Ay(z)=Ax(z), on the basis of the probability that differences
and Ax(z) arise solely from sampling error. If the differences
between
and Ax(z) are so large as to be unlikely, standard measures of
between
fit suggest that the model fits the data poorly, and vice versa if the differences
and Ax(z) are not so large as to be unlikely. The key point is
between
and Ax(z) are judged by how informative
that the differences between
the sample is about the population moments of yt. This is a sensible
procedure for judging the coherence of a null hypothesis, Ay(z)=Ax(z),
with the data. It is arguably less sensible when this null hypothesis is
known to be false.
Rather than rely on sampling error, the measures of fit that are developed
here are based on the size of the stochastic error required to reconcile the
autocovariances of xt with those of yt. In particular, let u, denote an n×1
error vector; then the importance of a difference between Ax(z) and Ay(z) will
be determined by asking, How much error would have to be added to xt so
that the autocovariances of xt+ut are equal to the autocovariances of yt? If
the variance of the required error is large, then the discrepancy between
Ax(z) and Ay(z) is large, and conversely if the variance of ut is small. The
vector ut is the approximation error in the economic model interpreted as a
stochastic process. It captures the second-moment characteristics of the
observed data that are not captured by the model. Loosely speaking, it is
analogous to the error term in a regression in which the set of regressors is
interpreted as the economic model. The economic model might be deemed
a good approximation to the data if the error term is small (i.e., the R2 of the
regression is large) and might be deemed a poor approximation if the error
term is large (i.e., the R2 of the regression is small).
To be more precise, assume that xt and yt are jointly covariance stationary,
and define the error ut by the equation
(1)
so that
(2)
where Au(z) is the ACGF of ut, Axy(z) is the cross ACGF between xt and yt,
and so forth. From the right-hand side of (2), three terms are needed to
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
MEASURES OF FIT
1016
307
JOURNAL OF POLITICAL ECONOMY
calculate Au(z). The first, Ay(z), can be consistently estimated from sample
data; the second, Ax(z), is completely determined by the model; but the
third, Axy(z), is not determined by the model, and it cannot be estimated
from the data (since this would require a sample drawn from the joint (xt,
yt) process). To proceed, an assumption is necessary.
A common assumption used in econometric analysis is that Axy(z) =Ax(z)
so that xt and ut are uncorrelated at all leads and lags. Equation (1) can then
be interpreted as the dynamic analogue of the classical errors-in-variables
model. Sargent (1989) discusses this assumption and an alternative
assumption, Axy(z)=Ay(z). He points out that under this latter assumption, ut
can be interpreted as signal extraction error, with yt an optimal estimate of
the unobserved “signal” xt.1 In many applications, these covariance restrictions
follow from the way the data were collected or the way expectations are
formed. For example, if x t represented the true value of the U.S.
unemployment rate and yt the value published by the U.S. Department of
Labor, then yt would differ from xt because of the sampling error inherent
in the monthly Current Population Survey from which yt is derived. The
sample design underlying the survey implies that the error term, ut, is
statistically independent of xt. Similarly, if yt denoted a rational expectation
of xt, then the error would be uncorrelated with yt. Neither of these
assumptions seems appropriate in the present context. The error is not the
result of imprecise measurement. It is not a forecast or signal extraction
error. Rather, it represents approximation or abstraction error in the economic
model. Any restriction used to identify Axy(z), and hence Au(z), is arbitrary.2
It is possible, however, to calculate a lower bound for the variance of ut
without imposing any restrictions on Axy(z). When this lower bound on
the variance of ut is large, then under any assumption on Axy(z), the model
fits the data poorly. If the lower bound on the variance of ut is small, then
there are possible assumptions about Axy(z) that imply that the model fits
the data well. Thus the bound is potentially useful for rejecting models
1
The reader familiar with work on data revisions will recognize these two sets of
assumptions as the ones underlying the “news” and “noise” models of Mankiw,
Runkle, and Shapiro (1984) and Mankiw and Shapiro (1986).
2
It is interesting to note that it is possible to determine whether the dynamic
errors-in-variables model or the signal extraction error model is consistent with the
model and the data. The dynamic errors-in-variables model implies that Ay(z)-Ax(z)≥0
for
so that the spectrum of yt lies everywhere above the spectrum of xt; the
signal extraction error model implies the converse. If the spectrum of x, lies anywhere
above the spectrum of yt, the errors-in-variables model is inappropriate; if the spectrum of yt lies anywhere above the spectrum of xt, the signal extraction model is
inappropriate. If the spectra of xt and yt cross, neither model is appropriate.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
308
CELEBRATION
MEASURES OF FIT
1017
on the basis of their empirical fit. Needless to say, models that appear to fit
the data well using this bound require further scrutiny.
The bound is calculated by choosing Axy(z) to minimize the variance of
ut subject to the constraint that the implied joint ACGF for xt and yt is
positive semidefinite. Equivalently, since the spectrum is proportional to
the ACGF evaluated at z=e-iω, the cross spectrum between xt and yt, (2π)1
Axy(e-iω), must be chosen so that the spectral density matrix of is
positive semidefinite at all frequencies.
Since the measures of fit proposed in this paper are based on the solution
to this minimization problem, it is useful to discuss the problem and its
solution in detail. Rather than move directly to the solution of the general
problem, we shall first solve two simpler problems. This helps develop
intuition for the general solution. In the first problem, xt and yt are serially
uncorrelated scalars, and the representation follows by inspection. In the
second problem, xt and yt are serially uncorrelated n×1 vectors, and the
solution is slightly more difficult to derive. Finally, in the last problem, xt
and yt are allowed to be serially correlated.
Model 1
Suppose that xt, yt, and ut are scalar serially uncorrelated random variables.
The problem is to choose σ xy to minimize the variance of u t ,
subject to the constraint that the covariance matrix
of xt and yt remains positive semidefinite, that is
By inspection,
the solution sets σxy=σxσy and yields
as the minimum. Since
σxy=σxσy, xt and yt are perfectly correlated with
(3)
where γ=σx/σy. Equation (3) is important because it shows how to
calculate fitted values of xt, given data on yt. Variants of equation (3)
will hold for all the models considered. In each model, the minimum
approximation error representation makes {xt} perfectly correlated with
{y t}. In each model, the analogue of (3) provides a formula for
calculating the fitted values of the variables in the model, given data
from the actual economy.
Model 2
Now suppose that xt and yt are serially uncorrelated random vectors with
covariance matrices Σx and Σy, respectively. Let Σu=Σx+Σy -Σxy-Σyx denote
the covariance matrix of ut. Since Σu is a matrix, there is not a unique way
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
MEASURES OF FIT
1018
309
JOURNAL OF POLITICAL ECONOMY
to judge how “small” it is. A convenient measure of the size of ut is the
where Σu,ij denotes the ijth element of Σu.
trace of Σu,
While convenient, this measure is not always ideal since it weights all
variables equally. Below, we shall find a representation that minimizes
tr(WΣu), where W is a prespecified n×n matrix. When all variables are
equally important, W=In, and unequal weighting can be implemented by
making W diagonal with the desired weights as the diagonal elements.
The matrix W can also be used to focus attention on specific linear
combinations of the variables that may be particularly interesting. For
example, let G denote an n×n matrix and suppose that the researcher is
primarily interested in the variables Gx t and Gy t . Then since
tr(GΣuG’)=tr(G’GΣu), W can be chosen as G’G.
The problem then is to choose Σxy to minimize tr(WΣu) subject to the
is positive semidefinite.
constraint that the covariance matrix of
The solution is given below for the case in which Σx has rank k≤n. This
occurs, for example, in economic models in which the number of variables
exceeds the number of shocks. The solution is summarized in the following
proposition.
PROPOSITION. Assume (i) rank(Σx)=k≤n, (ii) rank(WΣxW’)=
rank(Σx), and (iii) rank(Σy)=n. Let Cy denote an arbitrary n×n matrix square
and let Cx denote an arbitrary n×k matrix
root of Σy
Let USV’ denote the singular value
square root of Σx
where U is an n×k orthogonal matrix (U’U=Ik),
decomposition of
S is a k×k diagonal matrix, and V is a k×k orthonormal matrix. Then
is the unique matrix that minimizes tr(WΣu) subject to
is positive semidefinite.
the constraint that the covariance matrix of
The proof is given in the Appendix.
One important implication of this solution is that, like the scalar example,
is singular and xt can be represented as
the joint covariance matrix
(4)
where
(Since U’U=W=Ik, this simplifies to the scalar
result when xt and yt are scalars.)
Model 3
This same approach can be used in a dynamic multivariate model with
slight modifications; when ut is serially correlated, the weighted trace of
the spectral density matrix rather than the covariance matrix can be
minimized.
To motivate the approach, it is useful to use the Cramer representations
for xt, yt, and ut (see, e.g., Brillinger 1981, sec. 4.6). Assume that xt, yt,
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
310
CELEBRATION
MEASURES OF FIT
1019
and ut are jointly covariance stationary with mean zero; the Cramer
representation can be written as
(5)
where dz(ω)=(dz x(ω)′ dzy(ω)′ dzu(ω)′)′ is a complex valued vector of
orthogonal increments, with
where δ(ω-γ) is the dirac delta and S(ω) is the spectral density matrix of
at frequency ω. Equation (5) represents xt, yt, and ut as the
integral (sum) of increments dzx(ω), dzy(ω), and dzu(ω), which are uncorrelated
across frequencies and have variances and covariances given by the spectra
and cross spectra of xt, yt, and ut. Since the spectra are proportional to the
is proportional to Ax(e-iω),
ACGFs evaluated at z= e-iω,
-iω
is proportional to Axy(e ), and so forth.
Now consider the problem of choosing Axy(z) to minimize the variance
of ut. Since ut can be written as the integral of the uncorrelated increments
dzu(ω), the variance of ut can be minimized by minimizing the variance of
dzu(ω) for each ω. Since the increments are uncorrelated across frequency,
the minimization problems can be solved independently for each frequency.
Thus the analysis carried out for model 2 carries over directly, with spectral
density matrices replacing covariance matrices. The minimum trace
problem for model 2 is now solved frequency by frequency using the
spectral density matrix.
Like models 1–2, the solution yields
(6)
where Γ(ω) is the complex analogue of Γ from (4). Equation (6) implies
(7)
and
(8)
The autocovariances of ut follow directly from (8). Moreover, since dzx(ω)
and dzy(ω) are perfectly correlated from (7), xt can be represented as a
function of leads and lags of yt:
(9)
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
MEASURES OF FIT
1020
311
JOURNAL OF POLITICAL ECONOMY
where
with
xt can be calculated from leads and lags of yt.
Thus fitted values of
An Example
The model considered in the next section describes the dynamic properties of
output, consumption, investment, and employment as functions of a single
productivity shock. To demonstrate the mechanics of the minimum
approximation error representation for that model, assume that xt and yt are
n×1 vectors and that xt is driven by a single iid(0, 1) shock ∊t:
(10)
where α(L) is an n×1 matrix polynomial in the lag operator, L. Assume
that the Wold representation for the data is given by
(11)
where ⌰(L) is an n×n matrix polynomial in L, and et is an n×1 serially
uncorrelated vector with mean zero and identity covariance matrix.
The minimum error representation can then be computed directly from
the matrix expressions given in the proposition. From (10), Ax(z) =α(z)α(zl
)’ and, from (11), Αy(z)=⌰(z)⌰(z-1)’. Suppose that the weighting matrix is
W=In, so that the trace of the spectral density of ut is to be minimized for
α(e-iω)
each frequency. In terms of the matrices in the proposition, Cx(ω)=α
-iω
and Cy(ω)=⌰(e ). Thus the cross spectrum/cross ACGF for xt and yt is
chosen as Axy(e-iω)= α(e-iω)V(ω)U(ω)’⌰(eiω)’, where U(ω)S(ω)V(ω)’ is the
␣ (e-iω). (Since U(ω) and V(ω) are
singular value decomposition of ⌰(eiω)’␣
complex matrices, V(ω)’ and U(ω)’ denote the transpose conjugates of
V(ω) and U(ω), respectively.) The ACGF for ut follows from Au(e-iω)=
Ax(e-iω)+Ay(e-iω)-Axy(e-iω)-Ayx(e-iω). Finally, to compute fitted values of xt from
α (e the y t realization, note that dz x(ω)=Γ(ω)dz y(ω), where Γ(ω)=α
iω
-iω -1
)V(ω)U(ω)’⌰(e ) .
Relative Mean Square Approximation Error
A bound on the relative mean square approximation error for the economic
model can be calculated directly from (8). The bound—analogous to a lower
bound on 1-R2 from a regression—is
(12)
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
312
CELEBRATION
MEASURES OF FIT
1021
where [Au(z)]jj; and [Ay(z)]jj are the jth diagonal elements of Au(z) and Ay(z),
respectively. Thus rj(ω) is the variance of the jth component of dzu(ω)
relative to the jth component of dzy(ω), that is, the variance of the error
relative to the variance of the data for each frequency. A plot of rj(ω)
against frequency shows how well the economic model fits the data over
different frequencies. Integrating the numerator and denominator of rj(ω)
provides an overall measure of fit. (Note that since ut and xt are correlated,
rj(ω) can be larger than one; i.e., the R2 of the model can be negative.)3
One advantage of rj(ω) is that it is unaffected by time-invariant linear
filters applied to the variables. Filtering merely multiplies both the
numerator and denominator of rj(ω) by the same constant, the squared
gain of the filter. So, for example, rj(ω) is invariant to “Hod-rick-Prescott”
filtering (see Hodrick and Prescott 1980; King and Rebelo 1993) or
standard linear seasonal adjustment filters.4 The integrated version of the
relative mean square approximation error is not invariant to filtering since
it is a ratio of averages of both the numerator and denominator across
frequencies. When the data are filtered, the integrated version of rj(ω)
changes because the weights implicit in the averaging change. Frequencies
for which the filter has a large gain are weighted more heavily than
frequencies with a small gain.
III. Measures of Fit for a Real Business Cycle
Model
In this section, a standard real business cycle model is evaluated using
the measures of fit developed in the last section. The model, which derives
from Kydland and Prescott (1982), is the “baseline” model of King, Plosser,
and Rebelo (1988b). It is a one-sector neoclassical growth model driven
by an exogenous stochastic trend in technology.5
3
The measure rj(ω) is not technically a metric since it does not satisfy the triangle
inequality.
4
Standard seasonal adjustment filters such as the linear approximation to Census
X-11 have zeros at the seasonal frequencies, so that rj(ω) is undefined at these
frequencies for the filtered data.
5
This model is broadly similar to the model analyzed in Kydland and Prescott
(1982). While the baseline model does not include the complications of time to build,
inventories, time-nonseparable utility, and a transitory component to technology
contained in the original Kydland and Prescott model, these complications have been
shown to be reasonably unimportant for the empirical predictions of the model (see
Hansen 1985). Moreover, the King, Plosser, and Rebelo baseline model appears to fit
the data better at the very low frequencies than the original Kydland and Prescott
model since it incorporates a stochastic trend rather than the deterministic trend
present in the Kydland and Prescott formulation.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
MEASURES OF FIT
1022
313
JOURNAL OF POLITICAL ECONOMY
This baseline model is analyzed, rather than a more complicated variant,
for several reasons. First, the calibration/simulation exercises reported in
King, Plosser, and Rebelo suggest that the model explains the relative
variability of aggregate output, consumption, and investment, and it
produces series with serial correlation properties broadly similar to the
serial correlation properties of postwar U.S. data. Second, King, Plosser,
Stock, and Watson (1991) show that the low-frequency/cointegration
implications of the model are broadly consistent with similar postwar
U.S. data. Finally, an understanding of where this baseline model fits the
data and where it does not fit may suggest how the model should be
modified.
Only a brief sketch of the model is presented; a thorough discussion is
contained in King, Plosser, and Rebelo (1988a, 1988b). The details of the
model are as follows:
where Ct denotes consumption, Lt is leisure, Qt is output, Kt is capital, Nt is
employment, It is investment, and At is the stock of technology, with
log(At) assumed to follow a random walk with drift γa and innovation εt.
To analyze the model’s empirical predictions, the equilibrium of the
model must be calculated as a function of the parameters β, θ, α, γa,
and δ. This equilibrium implies a stochastic process for the variables Ct,
Lt, Nt, Kt, It, and Qt, and these stochastic processes can then be compared
to the stochastic processes characterizing U.S. postwar data. As is well
known, the equilibrium can be calculated by maximizing the representative
agent’s utility function subject to the technology and the resource
constraints. In general, a closed-form expression for the equilibrium does
not exist, and numerical methods must be used to calculate the stochastic
process for the variables corresponding to the equilibrium. A variety of
numerical approximations have been proposed (see Taylor and Uhlig [1990]
for a survey); here I use the log linearization of the Euler equations proposed
by King, Plosser, and Rebelo (1987). A formal justification for approximating
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
314
CELEBRATION
MEASURES OF FIT
1023
the equilibrium of this stochastic nonlinear model near its deterministic steady
state using linear methods is provided in Woodford (1986, theorem 2).
The approximate solution yields a vector autoregression (VAR) for the
logarithms of Qt,Ct,Kt, It, and Nt. (As in the standard convention, these
logarithms will be denoted by lowercase letters.) All of the variables except
nt are nonstationary but can be represented as stationary deviations about
at, the logarithm of the stock of technology, which by assumption follows
an integrated process. Thus qt, ct, it, and kt are cointegrated with a single
common trend, a t. Indeed, not only are the variables in the VAR
cointegrated, they are singular; the singularity follows since εt is the only
shock to the system. The coefficients in the VAR are complicated functions
and δ. Values for these
of the structural parameters β, θ, α, γa,
parameters are the same as those used by King, Plosser, and Rebelo (1988b):
when the variables are measured quarterly, the parameter values are α=.58,
δ=.025, γa= .004, σε=.010, and β=.988, and θ is chosen so that the steadystate value of N is .20. These parameter values were chosen so that the
model’s steady-state behavior matches postwar U.S. data.6 With these
values for the parameters, the VAR describing the equilibrium can be
calculated and the ACGF of xt=(Δqt Δct Δit nt)’ follows directly.7
These autocovariances will be compared to the autocovariances of
postwar data for the United States. The data used here are the same
data used by King, Plosser, Stock, and Watson (1991). The output
measure is total real private GNP, defined as total real GNP less
government purchases of goods and services. The measure of
consumption is total real consumption expenditures, and the measure
of investment is total real fixed investment. The measure of employment
is total labor hours in private nonagricultural establishments. All variables
are expressed in per capita terms using the total civilian noninstitutional
6
The choice of parameter values is described in King, Plosser, and Rebelo (1988a).
The value of a was chosen to equal the average value of labor’s share of gross national
product over 1948–86. The value of γa was chosen as the common average quarterly
rate of growth of per capita values of real GNP, consumption of nondurables and
services, and gross fixed investment. The depreciation rate was chosen to yield a
gross investment share of GNP of approximately 30 percent. The parameter θ was
chosen so that the model’s steady-state value of N matched the average workweek as
a fraction of total hours over 1948–86. The discaount rate β was chosen so that the
model’s steady-state annual interest rate matched the average rate of return on equity
over 1948–81. The value of σε=.01 appears to have been chosen as a convenient
normalization. This value is used here because it does a remarkably good job matching
the very low frequency movements in output, consumption, and investment.
7
Of course, this not the only possible definition of xt. The only restriction on xt is
covariance stationarity, so, e.g., ct-qt and it-qt could be included as elements.
8
All data are taken from Citibase. With the Citibase labels, the precise variables used
were gnp82—gge82 for output, gc82 for consumption, and gif82 for investment. The
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
MEASURES OF FIT
1024
315
JOURNAL OF POLITICAL ECONOMY
population over the age of 16.8 Let denote the logarithm of per capita
private output, the logarithm of per capita consumption expenditures,
and so forth. Then the data used in the analysis can be written as
The analysis presented in the last section assumed that the ACGF/
spectrum of yt was known. In practice, of course, this is not the case, and
the spectrum must be estimated. In this work, the spectrum of yt was
estimated in two different ways. First, an autoregressive spectral estimator
was used, calculated by first estimating a VAR for the variables and then
forming the implied spectral density matrix. As in King, Plosser, Stock,
and Watson (1991), the VAR was estimated imposing a cointegration
constraint between output, consumption, and investment. Thus the VAR
onto a
was specified as the regression of
constant and six lags of wt. The parameters of the VAR were estimated
using data for 1950–88. (Values before 1950 were used as lags in the
regression for the initial observations.) Second, a standard nonparametric
spectral estimator was also calculated. The spectrum was estimated by a
simple average of 10 periodogram ordinates after prewhitening employment
with the filter 1—.95L. These two estimators yielded similar values for the
measures of fit, and to conserve space only the results for the autoregressive
spectral estimator are reported.
For each variable, figure 1 presents the spectrum implied by the model,
the spectrum of the data, and the spectrum of the error required to reconcile
the model with the data.9 The error process was chosen to minimize the
unweighted trace of the error spectral density matrix, subject to the positive
semidefiniteness constraint discussed in the last section. Thus the objective
function weighted all the variables equally. For output, consumption,
and investment, the model and data spectra differ little for very low
frequencies (periods greater than 50 quarters) and for output and investment
at high frequencies (periods less than five quarters). There are significant
differences between the model and data spectra for periods typically
associated with the business cycle; the largest differences occur at a
frequency corresponding to approximately 20 quarters. The spectra of Δnt
and
are quite different. In addition to large differences at business
cycle frequencies, the spectra are also very different at low frequencies.
measure of total labor hours was constructed as total employment in nonagricultural
establishments (lhem) less total government employment (lpgov) multiplied by average
weekly hours (lhch). The population series was P16.
9
Figure 1 is reminiscent of figures in Howrey (1971, 1972), who calculated the
spectra implied by the Klein-Goldberger and Wharton models. A similar exercise is
carried out in Soderlind (1993), who compares the spectra of variables in the KydlandPrescott model to the spectra of postwar U.S. data.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
316
CELEBRATION
MEASURES OF FIT
1025
The model implies that employment is stationary so that its growth rate
has a spectrum that vanishes at frequency zero. In contrast, the data suggest
significant low-frequency variation in postwar U.S. employment.10
The figure shows that relatively little error is needed to reconcile the
model and the data for output, consumption, and investment over the
very low frequencies. On the other hand, error with a variance on the
order of 40–50 percent of the magnitude of the variance of the series is
necessary for the components of output, consumption, and investment
with periods in the 6–32-quarter range. At higher frequencies, the model
is able to match the stochastic process describing investment, but not the
processes describing the other series.
Table 1 provides a summary of the relative mean square approximation
error (RMSAE) for a variety of weighting functions and filters. Each
panel shows the RMSAE for the variables for five different minimum
error representations. Column 1 presents results for the representation
that obtains when the unweighted trace of the spectrum is minimized;
this is the representation used to construct the error spectra shown in
figure 1. Column 2 summarizes the results for the representation that
minimizes the output error, with no weight placed on the other variables.
Column 3 summarizes results for the representation that minimizes the
consumption error, and so forth. Panel A presents the results for the
differences of the data integrated across all frequencies, panel B shows
results for the levels of the series detrended by the Hodrick-Prescott filter
integrated across all frequencies, and panel C presents results for the
levels of the series integrated over business cycle frequencies (6–32 quarters).
The trade-off inherent in the different representations is evident in all
panels. For example, in panel A, with the minimum output error
representation, the RMSAE for output is 26 percent, and the RMSAE for
consumption is 78 percent; when the minimum consumption error
representation is chosen, the RMSAE for consumption falls to 21 percent
but the RMSAE for output rises to 75 percent. When all the variables are
equally weighted, the RMSAE is 52 percent for output and 66 percent for
consumption. Panel C shows that most of this trade-off occurs at the high
frequencies, at least for output, consumption, and investment; over the
business cycle frequencies their RMSAEs are in the 40–60 percent range.
As explained in Section II, fitted values of the model’s variables can be
constructed using the minimum error representation together with the
10
The figures do not include standard errors for the spectra estimated from these
data. These standard errors are large—approximately one-third the size of the estimated
spectra. The standard errors for the RMSAE, averaged across frequencies, are
considerably smaller. These are included in tables 1 and 2 below.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
MEASURES OF FIT
317
FIG. 1.—Decomposition of spectra: a, output; b, consumption; c, investment; d, employment. Dotted lines refer to the data, dashed lines to the model, and solid lines to approximation error.
1026
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
318
CELEBRATION
FIG. 1.—Continued
1027
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
TAB LE 1
RELATIVE MEAN SQUARE APPROXIMATION ERROR
NOTE.—Relative mean square approximation error is the lower bound of the variance of the approximation error divided by the variance of the series. Each column represents the relative mean
square approximation error of the row variable constructed from the representation that minimizes the weighted trace of the error spectrum. The weights are summarized by the column headings.
For example, col. 1 is the equally weighted trace, col. 2 puts all the weight on the output error, etc. The numbers in parentheses are standard errors based on the sampling error in the estimated VAR
coefficients used to estimate the data spectrum.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
320
CELEBRATION
MEASURES OF FIT
1029
observed data. Since the measurement error model represents yt as xt plus
error, the standard signal extraction formula can be used to extract {xt}
from {yt}. In general, of course, signal extraction methods will yield an
that is not exact in the sense that
estimate of xt, say
In the present context, the estimate will be exact since the measurement
error process is chosen so that dzx(ω) and dzy(ω) are perfectly correlated for
all ω.11 Figure 2 shows the realizations of the data and the realizations of
the variables in the model calculated from the data using the equally
weighted minimum output error representation.12
In figure 2a, which shows the results for output, the model seems
capable of capturing the long swings in the postwar U.S. data but not
capable of capturing the cyclical variability in the data. Private per capita
GNP fell by 8.4 percent from the cyclical peak in 1973 to the trough in
1975 and by 6.8 percent from the peak in 1979 to the trough in 1982. In
contrast, the corresponding drops in Qt—output in the model—were 3.1
percent and 3.0 percent, respectively. The dampened cyclical swings in
consumption and investment, shown in figure 2b and c, are even more
dramatic. Finally, figure 2d shows that the model predicts changes in
employment that have little to do with the changes observed in the United
States during the postwar period.13
One possible explanation for the relatively poor fit of the model is that
the “calibrated” values of the parameters are wrong. In particular,
Christiano and Eichenbaum (1990) show that the model’s predictions
change in an important way when the technology process changes from a
random walk to a stationary AR(1). Table 2 shows how the model fares
for a range of values of the AR(1) coefficient for technology, denoted by
ρa. Panel A of the table shows the results for first differences of the variables
across all frequencies, panel B presents results for the Hodrick-Prescott
detrended levels of the series, and panel C shows the results for the levels
11
More precisely, the estimate is exact in the sense that
converges in mean square to xt as j→∞.
12
As shown in eq. (9), xt can be calculated as β(L)yt, where β(L) is the inverse
Fourier transform of G(ω). To calculate the estimates shown in the figure, Γ(ω) was
calculated at 128 equally spaced frequencies between zero and π. Since β(L) is twosided, pre-and postsample values of yt are required to form β(L)xt. These pre- and
postsample values were replaced with the sample means of the yt data. The first
differences, xt and yt, were then accumulated to form the levels series shown in the
figure.
13
The calculations required to construct figs. 1 and 2 and the results in table 1 are
easily carried out. For this example, the model spectrum, data spectrum, RMSAEs,
and fitted values were calculated in less than a minute on a standard desktop computer.
A GAUSS program for these calculations is available from the author.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
MEASURES OF FIT
321
FIG. 2.—Data: a, output; b, consumption; c, investment; d, employment. Solid lines refer
to U.S. data and dashed lines to realizations from the model.
1030
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
322
CELEBRATION
FIG. 2—Continued
1031
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
MEASURES OF FIT
1032
323
JOURNAL OF POLITICAL ECONOMY
TAB LE 2
RELATIVE MEAN SQUARE APPROXIMATION ERROR AS A FUNCTION OF
THE AR(1) COEFFICIENT FOR TECHNOLOGY
NOTE.—Relative mean square approximation error is the lower bound of the variance of the approximation
error divided by the variance of the series. Each column represents the relative mean square approximation error of
the row variable constructed from the representation that minimizes the weighted trace of the error spectrum. The
column headings represent the AR(1) coefficient for the process of the logarithm of productivity in the model. For
example, col. 1 represents results for the model with random walk technological progress. The numbers in parentheses
are standard errors based on the sampling error in the estimated VAR coefficients used to estimate the data spectrum.
of the series over the “business cycle” frequencies. From panel A, the
value of ρa has little effect on the fit of the model averaged across all
frequencies. In particular, as ρα falls from 1.0 to .90, the RMSAE increases
slightly for consumption, falls for investment, and changes little for output
and employment. In contrast, the value of ρa has a significant effect on the
fit of the model over business cycle frequencies. For example, panel C
shows that as ρa falls from 1.0 to .90, the RMSAE falls for output (.43 to
.33), for investment (.42 to .17), and for employment (.72 to .52); it increases
for consumption (.52 to .66).
The source of the changes in the RMSAEs can be seen in figure 3,
which plots the spectra of the variables in models with ρa=1 and ρa=.90.
The largest difference between the spectra of the models is the increase
in variance in output, investment, and employment as ρa falls from 1.0 to
ρa.90. The economic mechanism behind this increased variance is the
increase in intertemporal substitution in response to a technology shock.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
324
CELEBRATION
MEASURES OF FIT
1033
When ρa=.90, technology shocks are transitory, and they can be exploited
only by large transitory increases in employment and investment. It is
interesting to note that while this mechanism increases the variance of the
growth rates in employment, investment, and output, it has little effect on
their autocorrelations. That is, as ρα changes from 1.0 to .90, the shape of
the spectra changes little.
Before we leave this section, six additional points deserve mention.
First, the fitted values in figure 2 are quantitatively and conceptually
similar to figures presented in Christiano (1988) and Plosser (1989). They
calculated the Solow residual from actual data and then simulated the
economic model using this residual as the forcing process. Implicitly,
they assumed that the model and data were the same in the terms of their
Solow residual, and then asked whether the model and data were similar
in other dimensions. Figure 2 is constructed by making the model and
data as close as possible in one dimension (in this case the trace of the
variance of the implied approximation error) and then asking whether the
model and data are similar in other dimensions. The difference between
the two approaches can be highlighted by considering the circumstances
in which they would produce exactly the same figure. If the Solow residual
computed from the actual data followed exactly the same stochastic process
as the change in productivity in the model, and if the approximation
error representation was constructed by minimizing the variance of the
difference between the Solow residual in the data and productivity growth
in the model, then the two figures would be identical. Thus the figures
will differ if the stochastic process for the empirical Solow residual is not
the same as assumed in the model, or the approximation error representation
is chosen to make the model and data close in some dimension other than
productivity growth.
Second, the inability of the model to capture the business cycle properties
of the data is not an artifact of the minimum measurement error
representation used to form the projection of xt onto yτ, τ = 1, …, n.
Rather, it follows directly from a comparison of the spectra of xt and yt.
The fitted values are constrained to have an ACGF/ spectra given by the
economic model. Figure 1 shows that, for all the variables, the spectral
power over the business cycle frequencies is significantly less for the model
than for the data. Therefore, fitted values from the model are constrained
to have less cyclical variability than the data.
Third, the ability of the model to mimic the behavior of the data depends
critically on the size of the variance of the technology shock. The value of
σε used in the analysis is two and one-half times larger than the drift in the
series. Thus if the εt were approximately normally distributed, the stock of
technology At would fall in one out of three quarters on average. Reducing
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
MEASURES OF FIT
325
FIG. 3.—Data and model spectra: a, output; b, consumption; c, investment; d, employment. Dotted lines refer to the data, solid lines to the model with ρα=1.00, and dashed lines
to the model with ρα=.90.
1034
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
326
CELEBRATION
FIG.3—Continued
1035
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
MEASURES OF FIT
1036
327
JOURNAL OF POLITICAL ECONOMY
the standard deviation of the technology shock so that it equals the average
growth in at drastically increases the size of the measurement error necessary
to reconcile the model with the data. For example, integrated across all
frequencies, the RMSAE for output increases from 52 percent to 74 percent.
Fourth, there is nothing inherent in the structure of the model that
precludes the use of classical statistical procedures. Altug (1989) used
maximum likelihood methods to study a version of the model that is
augmented with serially correlated classical measurement errors. Singleton
(1988) and Christiano and Eichenbaum (1992) pointed out that generalized
method of moments procedures can be used to analyze moment implications
of models like the one presented above. In the empirical work of Christiano
and Eichenbaum the singularity in the probability density function of the
data that is implied by the model was finessed in two ways. First, limited
information estimation and testing methods were used, and second, the
authors assumed that their data on employment were measured with error.
Fifth, many if not all of the empirical shortcomings of this model have
been noted by other researchers. King, Plosser, and Rebelo clearly show
that the model is not capable of explaining the variation in employment
that is observed in the actual data. The implausibility of the large
technology shocks is discussed in detail in Summers (1986), Mankiw
(1989), and McCallum (1989).
Finally, the analysis above has concentrated on the ability of the model
to explain the variability in output, consumption, investment, and
employment across different frequencies. While it is possible to analyze
the covariation of these series using the cross spectrum of the measurement
error, such an analysis has not been carried out here. This is a particularly
important omission since this is the dimension in which the baseline real
business cycle model is typically thought to fail. For example, Christiano
and Eichenbaum (1992) and Rotemberg and Woodford (1992) use the
model’s counterfactual implication of a high correlation between average
productivity and output growth as starting points for their analysis, and
the empirical literature on the intertemporal capital asset pricing model
beginning with Hansen and Singleton (1982) suggests that the asset pricing
implications of the model are inconsistent with the data. It would be
useful to derive simple summary statistics based on the cross spectra of the
measurement error and the data to highlight the ability of the model to
explain covariation among the series.
IV. Discussion
The discussion thus far has assumed that the parameter values of the
economic model are known. A natural question is whether the measures
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
328
CELEBRATION
MEASURES OF FIT
1037
of fit discussed in this paper can form the basis for estimators of these
parameters. Does it make sense, for example, to estimate unknown
parameters by minimizing some function of the relative mean square error,
rj(ω) given in equation (12)? This certainly seems sensible. For example, a
researcher may want to “calibrate” his model with a value ofρα=.90 rather
than 1.0, because this value produces spectra closer to the estimated spectra
of data over the business cycle frequencies. Yet dropping the standard
statistical assumption that the economic model is correctly specified raises
a number of important issues. Foremost among these is the meaning of
the parameters. If the model does not necessarily describe the data, then
what do the parameters measure? Presumably, the model is meant to
describe certain characteristics of the data’s stochastic process (the business
cycle or the growth properties, for example), while ignoring other
characteristics. It then makes sense to define the model’s parameters as
those that minimize the differences between the model and the data’s
stochastic process in dimensions that the model is attempting to explain.
So, for example, it seems sensible to define the parameters of a growth
model as those that minimize rj(ω) over very low frequencies, or to define
the parameters of a business cycle model as those that minimize rj(ω) over
business cycle frequencies. Given this definition of the parameters,
constructing an analog estimator (see Manski 1988) by minimizing
corresponds to a standard statistical practice.
Of course, the parameters may also be defined using other characteristics
of the model and the stochastic process describing the data. For example,
in standard “calibration” estimation exercises, many of the parameters are
implicitly defined in terms of first moments of the data. Parameters are
chosen so that the first moments of the variables in the model’s steady
state match the first moments of the data.
Two final points deserve mention. First, since the measures of fit
developed in this paper are based on a representation that minimizes the
discrepancy between the model and the data, they serve only as a bound
on the fit of the model. Models with large RMSAEs do not fit the data
well. Models with small RMSAEs fit the data well given certain
assumptions about the correlation properties of the noise that separates
the model and the data, but may fit the data poorly given other assumptions
about the noise.
Finally, while this paper has concentrated on measures of fit motivated
by a model of measurement error, other measures are certainly possible.
For example, one measure, which like the measures in this paper uses
only the autocovariances implied by the model and the data, is the expected
log likelihood ratio using the normal probability density function (pdf) of
the data and the model. More precisely, if g(x) denotes the normal pdf
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
MEASURES OF FIT
1038
329
JOURNAL OF POLITICAL ECONOMY
constructed from the autocovariances of the data, f(x) denotes the normal
pdf constructed for the autocovariances implied by the model, and Eg is
the expectation operator taken with respect to g(x), the expected log
likelihood ratio I(g, f)= Eg{log[g(x)/f(x)]} can be used to measure the distance
between the densities f(·) and g(·); I(g, f) is the Kullback-Leibler information
criterion (KLIC), which plays an important role in the statistical literature
on model selection (e.g., Akaike 1973) and quasi-maximum likelihood
estimation (White 1982). Unfortunately, the KLIC will not be defined
when f(x) is singular and g(x) is not; the KLIC distance between the two
densities is infinite. Thus, for example, it would add no additional
information on the fit of the real business cycle model analyzed in Section
III beyond pointing out the singularity.
Arguably, one of the most informative diagnostics presented in this
paper is the plot of the model and data spectra. For example, figures 1 and
2 show that the data spectra have mass concentrated around the business
cycle frequencies, but the model spectra do not. Any metric comparing
the data and model spectra may serve as a useful measure of fit. The
RMSAE proposed here has the advantage that it can be interpreted like 1R2 from a regression, but any summary statistic discards potentially useful
information contained in plots such as figures 1 and 2. Some practical
advice, therefore, is to present both model and data spectra as a convenient
way of comparing their complete set of second moments.
Appendix
To prove the proposition, first parameterize Σx, Σy, and Σxy as
(A1)
(A2)
(A3)
where Cx is n×k with full column rank, G is n×k, and Σ is positive semidefinite. Since
ΣU=Σx+Σy-Σxy-Σyx, minimizing tr(WΣu) with Σx and Σy given is equivalent to maximizing
tr(WΣxy)=tr(WCxG’). Given an arbitrary factorization of Σx of the form (A1), the problem is
to find the n×k matrix G to maximize tr(WCxG’) subject to the constraint that Σy- GG´=Σ is
positive semidefinite.
Then Σy-GG´ is positive semidefinite if and only if In´ is positive
Let
semidefinite, which in turn is true if and only if all the eigenvalues of GG’ are less than or
´ are the same as those of
, the problem can
equal to one. Since the eigenvalues of
be written as
(A4)
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
330
CELEBRATION
MEASURES OF FIT
where λ i(
) denotes the ith eigenvalue of
tr(AB)=tr(BA) for conformable matrices A and B.
1039
, and I have used the fact that
Let QDR´ denote the singular value decomposition of , where Q is an n ×k orthogonal
matrix, R is a k×k orthonormal matrix, and D is a k×k diagonal matrix with elements dij. Since
and
since
the solution to (A4) is seen to require that λi(
)=1, i=1, …, k. This
implies that
=I. Write the singular value decomposition of
as USV´; then
where
Since
the maximization problem can be written as
(A5)
Assumptions i-iii of the proposition imply that
has full column rank so that S is a
diagonal matrix with strictly positive diagonal elements. Thus since U´U=Ik, the maximization is achieved by
Working backward, we see that G=C y UV’, so that
Uniqueness follows since this choice of Σxy does not depend on the (arbitrary) choice of
the matrix square roots, Cx and Cy. To see this, let y and x denote other matrix square roots
of Σy and Σx. Then y=CyRy and x =CxRx, where Ry and Rx are orthonormal matrices.
From the analysis above, this yields
where
is the singular value
decomposition of
. By inspection,
and
so that
References
Akaike, H. “Information Theory and an Extension of the Maximum Likelihood Principle.”
In Proceedings of the Second International Symposium on Information Theory, edited by B.N.Petrov
and F.Csaki. Budapest: Akademiai Kaido, 1973.
Altug, Sumru. “Time-to-Build and Aggregate Fluctuations: Some New Evidence.” Internal.
Econ. Rev. 30 (November 1989): 889–920.
Brillinger, David R. Time Series: Data Analysis and Theory. San Francisco: Holden-Day, 1981.
Campbell, John Y., and Shiller, Robert J. “The Dividend-Price Ratio and Expectations of
Future Dividends and Discount Factors.” Rev. Financial Studies 1, no. 3 (1988): 195–228.
Christiano, Lawrence J. “Why Does Inventory Investment Fluctuate So Much?” J.Monetary
Econ. 21 (March/May 1988): 247–80.
Christiano, Lawrence J., and Eichenbaum, Martin. “Unit Roots in Real GNP: Do We Know
and Do We Care?” Carnegie-Rochester Conf. Ser. Public Policy 32 (Spring 1990): 7–61.
————. “Current Real-Business-Cycle Theories and Aggregate Labor-Market Fluctuations.”
A.E.R. 82 (June 1992): 430–50.
Cochrane, John H. “Explaining the Variance of Price-Dividend Ratios.” Rev. Financial Studies
5, no. 2 (1992): 243–80.
Durlauf, Steven N., and Hall, Robert E. “Measuring Noise in Stock Prices.” Manuscript.
Stanford, Calif.: Stanford Univ., 1989.
Geweke, John. “The Dynamic Factor Analysis of Economic Time-Series Models.” In Latent
Variables in Socio-economic Models, edited by Dennis J.Aigner and Arthur S.Goldberger.
Amsterdam: North-Holland, 1977.
Haavelmo, Trygve. “The Probability Approach in Econometrics.” Econometrica 12 (suppl.;
July 1944): 1–115.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
MEASURES OF FIT
1040
331
JOURNAL OF POLITICAL ECONOMY
Hansen, Gary D. “Indivisible Labor and the Business Cycle.” J.Monetary Econ. 16 (November
1985): 309–27.
Hansen, Gary D., and Sargent, Thomas J. “Straight Time and Overtime in Equilibrium.”
J.Monetary Econ. 21 (March/May 1988): 281–308.
Hansen, Lars Peter, and Jagannathan, Ravi. “Implications of Security Market Data for Models
of Dynamic Economies.” J.P.E. 99 (April 1991): 255–62.
Hansen, Lars Peter, and Sargent, Thomas J. “Formulating and Estimating Dynamic Linear
Rational Expectations Models.” J.Econ. Dynamics and Control 2 (February 1980): 7–46.
Hansen, Lars Peter, and Singleton, Kenneth J. “Generalized Instrumental Variables Estimation
of Nonlinear Rational Expectations Models.” Econometrica 50 (September 1982): 1269–86.
Hodrick, Robert J., and Prescott, Edward C. “Post-War U.S. Business Cycles: An Empirical
Investigation.” Manuscript. Pittsburgh: Carnegie Mellon Univ., 1980.
Howrey, E.Philip. “Stochastic Properties of the Klein-Goldberger Model.” Econometrica 39
(January 1971): 73–87.
——. “Dynamic Properties of a Condensed Version of the Wharton Model.” In Econometric
Models of Cyclical Behavior, vol. 2, edited by Bert G. Hickman. New York: Columbia Univ.
Press (for NBER), 1972.
King, Robert G.; Plosser, Charles I.; and Rebelo, Sergio T. “Production, Growth, and
Business Cycles: Technical Appendix.” Manuscript. Rochester, N.Y.: Univ. Rochester,
1987.
——. “Production, Growth and Business Cycles: I. The Basic Neoclassical Model.” J.Monetary
Econ. 21 (March/May 1988): 195–232. (a)
——. “Production, Growth and Business Cycles: II. New Directions.” J. Monetary Econ. 21
(March/May 1988): 309–41. (b)
King, Robert G.; Plosser, Charles I.; Stock, James H.; and Watson, Mark W. “Stochastic
Trends and Economic Fluctuations.” A.E.R. 81 (September 1991): 819–40.
King, Robert G., and Rebelo, Sergio T. “Low Frequency Filtering and Real Business Cycle
Models.” J.Econ. Dynamics and Control (1993), in press.
Kydland, Finn E., and Prescott, Edward C. “Time to Build and Aggregate Fluctuations.”
Econometrica 50 (November 1982): 1345–70.
McCallum, Bennett T. “Real Business Cycle Models.” In Modern Business Cycle Theory, edited
by Robert J.Barro. Cambridge, Mass.: Harvard Univ. Press, 1989.
McFadden, Daniel. “Econometric Models of Probabilistic Choice.” In Structural Analysis of
Discrete Data with Econometric Applications, edited by Charles Manski and Daniel McFadden.
Cambridge, Mass.: MIT Press, 1981.
Mankiw, N.Gregory. “Real Business Cycles: A New Keynesian Perspective.” J.Econ. Perspectives
3 (Summer 1989): 79–90.
Mankiw, N.Gregory; Runkle, David E.; and Shapiro, Matthew D. “Are Preliminary
Announcements of the Money Stock Rational Forecasts?” J.Monetary Econ. 14 (July 1984):
15–27.
Mankiw, N.Gregory, and Shapiro, Matthew D. “News or Noise: An Analysis of GNP Revisions.”
Survey Current Bus. 66 (May 1986): 20–25.
Manski, Charles F. Analog Estimation Methods in Econometrics. New York: Chapman and Hall,
1988.
Plosser, Charles I. “Understanding Real Business Cycle Models.” J.Econ. Perspectives 3 (Summer
1989): 51–77.
Prescott, Edward C. “Theory Ahead of Business-Cycle Measurement.” Carnegie-Rochester
Conf. Ser. Public Policy 25 (Autumn 1986): 11–44.
Rotemberg, Julio J., and Woodford, Michael. “Oligopolistic Pricing and the Effects of Aggregate
Demand on Economic Activity.” J.P.E. 100 (December 1992): 1153–1207.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
332
CELEBRATION
MEASURES OF FIT
1041
Sargent, Thomas J. “Two Models of Measurements and the Investment Accelerator.” J.P.E.
97 (April 1989): 251–87.
Sargent, Thomas J., and Sims, Christopher A. “Business Cycle Modeling without Pretending
to Have Too Much a Priori Economic Theory.” In New Methods in Business Cycle Research, by
Christopher A.Sims et al. Minneapolis: Fed. Reserve Bank, 1977.
Singleton, Kenneth J. “Econometric Issues in the Analysis of Equilibrium Business Cycle
Models.” J.Monetary Econ. 21 (March/May 1988): 361–86.
Soderlind, Paul. “Cyclical Properties of a Real Business Cycle Model.” Manuscript. Princeton,
N.J.: Princeton Univ., Dept. Econ., 1993.
Summers, Lawrence H. “Some Skeptical Observations on Real Business Cycle Theory.” Fed.
Reserve Bank Minneapolis Q. Rev. 10 (Fall 1986): 23–27.
Taylor, John B., and Uhlig, H. “Solving Nonlinear Stochastic Growth Models: A Comparison
of Alternative Solution Methods.” J.Bus. and Econ. Statis. 8 (January 1990): 1–17.
White, Halbert. “Maximum Likelihood Estimation of Misspecified Models.” Econometrica 50
(January 1982): 1–25.
Woodford, Michael. “Stationary Sunspot Equilibria: The Case of Small Fluctuations around
a Deterministic Steady State.” Manuscript. Chicago: Univ. Chicago, 1986.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
CHAPTER 18
333
JOURNAL OF APPLIED ECONOMETRICS, VOL. 9, S123–S144 (1994)
STATISTICAL INFERENCE IN CALIBRATED MODELS
FABIO CANOVA
Department of Economics, Universitat Pompeu Fabra, Balmes 132, 08008 Barcelona, Spain
and Department of Economics, Università di Catania, 95100 Catania, Italy, and CEPR
SUMMARY
This paper describes a Monte Carlo procedure to assess the performance of calibrated dynamic
general equilibrium models. The procedure formalizes the choice of parameters and the evaluation
of the model and provides an efficient way to conduct a sensitivity analysis for perturbations of the
parameters within a reasonable range. As an illustration the methodology is applied to two problems: the equity premium puzzle and how much of the variance of actual US output is explained by
a real business cycle model.
1. INTRODUCTION
The current macroeconometrics literature has proposed two ways to confront general
equilibrium rational expectations models with data. The first, an estimation approach, is the
direct descendant of the econometric methodology proposed 50 years ago by Haavelmo
(1944). The second, a calibration approach, finds its justification in the work of Frisch
(1933) and is closely linked to the computable general equilibrium literature surveyed e.g. in
Shoven and Whalley (1984).
The two methodologies share the same strategy in terms of model specification and
solution. Both approaches start from formulating a fully specified general equilibrium
dynamic model and in selecting convenient functional forms for preferences, technology,
and exogenous driving forces. They then proceed to find a decision rule for the
endogenous variables in terms of the exogenous and predetermined variables (the states)
and the parameters. When the model is nonlinear, closed-form expressions for the decision
rules may not exist and both approaches rely on recent advantages in numerical methods to
find an approximate solution which is valid either locally or globally (see e.g. the January
1990 issue of the Journal of Business and Economic Statistics for a survey of the methods and
Christiano, 1990, and Dotsey and Mao, 1991, for a comparison of the accuracy of the
approximations).
It is when it comes to choosing the parameters to be used in the simulations and in
evaluating the performance of the model that several differences emerge. The first
procedure attempts to find the parameters of the decision rule that best fit the data either by
maximum likelihood (ML) (see e.g. Hansen and Sargent, 1979, or Altug, 1989) or
generalized method of moments (GMM) (see e.g. Hansen and Singleton, 1983, or Burnside
et al., 1993). The validity of the specification is examined by testing restrictions, by general
goodness of fit tests or by comparing the fit of two nested models. The second approach
‘calibrates’ parameters using a set of alternative rules which includes matching long-run
averages, using previous microevidence or a priori selection, and assesses the fit of the model
with an informal distance criterion.
These differences are tightly linked to the questions the two approaches ask. Roughly
speaking, the estimation approach asks the question ‘Given that the model is true, how false
CCC 0883–7252/94/0S0S123–22
© 1994 by John Wiley & Sons, Ltd.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
Received July 1992
Revised August 1994
334
S124
CALIBRATION
F. CANOVA
is it?’ while the calibration approach asks ‘Given that the model is false, how true is it?’
Implicit in the process of estimation is in fact the belief that the probability structure of a
model is sufficiently well specified to provide an accurate description of the data. Because
economic models are built with an emphasis on tractability, they are often probabilistically
underspecified so that measurement errors or unobservable shocks are added at the
estimation stage to complete their probability structure (see e.g. Hansen and Sargent, 1980,
or Altug, 1989). By testing the model, a researcher takes the model seriously as a datagenerating process (DGP) and examines what features of the specification are at variance
with the data. A calibrationist takes the opposite view: the model, as a DGP for the data, is
false. That is, as the sample size grows, it is known that the data are generated by the model
will be at increasingly greater variance with the observed time series. An economic model is
seen, at best, as an approximation to the true DGP which need not be either accurate or
realistic and, as such, should not be regarded as a null hypothesis to be statistically tested
(see Prescott, 1991, p. 5). In confronting the model with the data, a calibrationist wants to
indicate the dimensions where the approximation is poor and suggest modifications to the
theoretical model in order to obtain a better approximation.
Both methodologies have weak points. Model estimation involves a degree of
arbitrariness in specifying which variables are unobservable or measured with error. In the
limit, since all variables are indeed measured with error, no estimation seems possible and
fruitful. In addition, tests of the model’s restrictions may fail to indicate how to alter the
specification to obtain a better fit. The limitations of the calibration approach are also well
known. First, the selection criterion for parameters which do not measure long-run
averages is informally specified and may lead to contradictory choices. Information used in
different studies may in fact be inconsistent (e.g. a parameter chosen to match labour
payments from firms in national account data may not equal the value chosen to match the
labour income received by households) and the range of estimates for certain parameters
(e.g. risk aversion parameter) is so large that selection biases may be important. Second, the
outcomes of the simulations typically depend on the choice of unmeasured parameters.
However, although some authors (see e.g. Prescott, 1991, p. 7, or Kydland, 1992, p. 478)
regard a calibration exercise as incomplete unless the sensitivity of the results to reasonable
perturbations of the parameters selected a priori or not well tied to the data is reported, such
an analysis is not often performed. Third, because the degree of confidence in the results
depends on both the degree of confidence in the theory and in the underlying measurement
of the parameters and because either parameter uncertainty is disregarded or, when a
search is undertaken, the number of replications typically performed is small, we must
resort to informal techniques to judge the relevance of the theory.
This paper attempts to eliminate some of the weaknesses of calibration procedures while
maintaining the general analytical strategy employed in calibration exercises. The focus is
on trying to formalize the selection of the parameters and the evaluation process and in
designing procedures for meaningful robustness analysis on the outcomes of the
simulations. The technique we propose shares features with those recently described by
Gregory and Smith (1991) and Kwan (1990), has similarities with stochastic simulation
techniques employed in dynamic nonlinear large scale macro models (see e.g. Fair, 1991),
and generalizes techniques on randomized design for strata existing in the static
computable general equilibrium literature (see e.g. Harrison and Vinod, 1989).
The idea of the technique is simple. We would like to reproduce features of actual data,
which is taken to be the realization of an unknown vector stochastic process, with an
‘artificial economy’ which is almost surely the incorrect generating mechanism for the
actual data. The features we may be interested in include conditional and unconditional
moments (or densities), the autocovariance function of the data, functions of these
quantities (e.g. measures of relative volatility), or specific events (e.g. a recession). A model
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
STATISTICAL INFERENCE IN CALIBRATED MODELS
STATISTICAL INFERENCE IN CALIBRATED MODELS
335
S125
is simulated repeatedly using a Monte Carlo procedure which randomizes over both the
exogenous stochastic processes and the parameters. Parameters are drawn from a data-based
density which is consistent with the information available to a simulator (which may
include both time-series and cross-sectional aspects). We judge the validity of a model on its
ability to reproduce a number of ‘stylized facts’ of the actual economy (see Friedman, 1959).
The metric used to evaluate the discrepancy of the model from the data is probabilistic. We
construct the simulated distribution of the statistics of interest and, taking the actual
realization of the statistic as a critical value, examine (1) in what percentile of the simulated
distribution the actual value lies and (2) how much of the simulated distribution is within
a k% region centred around the critical value. Extreme values for the percentile (say, below
α% or above (1-α)%) or a value smaller than k for the second probability indicates a poor fit
in the dimensions examined.
The approach has several appealing features. First, it accounts for the uncertainty faced by
a simulator in choosing the parameters of the model in a realistic way. Second, it has a built-in
feature for global sensitivity analysis on the support of the parameter space and allows for
other forms of conditional or local sensitivity analysis. Third, it provides general evaluation
criteria and a simple and convenient framework to judge the relevance of the theory.
The paper is divided into six sections. The next section introduces the technique,
provides a justification for the approach and describes the details involved in the
implementation of the procedure. Section 3 deals with robustness analysis. Section 4 spells
out the relationship with existing techniques. Two examples describing the potential of the
technique for problems of practical interest appear in Section 5. Section 6 presents
conclusions.
2. THE TECHNIQUE
A General Framework of Analysis
We assume that a researcher is faced with an m×1 vector of time series
which are the
realizations of a vector stochastic process and that she is interested in reproducing features
of
using a dynamic general equilibrium model.
is assumed to have a continuous but
unknown distribution and moments up to the nth. For the sake of presentation we assume
is independent of t but shifts in the unconditional
that the unconditional distribution of
distribution of at known points can easily be handled. may include macro variables like
GNP, consumption, interest rates, etc. We also assume that dynamic economic theory gives
us a model expressing the endogenous variables Xt as a function of exogenous and predetermined variables Zt (the states of the problem) and of the parameters β. Zt may include objects
like the existing capital stock, exogenous fiscal, and monetary variables or shocks to technologies and preferences. We express the model’s functional relation as Xt=f(Zt, β). Under specific
assumptions about the structure of the economy (e.g. log or quadratic preferences, CobbDouglas production function, full depreciation of the capital stock in the one-sector growth
model), f can be computed analytically either by value function iteration or by solving the
Euler equations of the model subject to the transversality condition (see e.g. Hansen and
Sargent, 1979). In general, however, f cannot be derived analytically from the primitives of the
problem. A large body of current literature has concentrated on the problem of finding
approximations which are either locally or globally close to f for a given metric.1
1
These include linear or log-linear expansions of f around the steady state (Kydland and Prescott, 1982; and
King et al., 1988), backward-solving methods (Sims, 1984; Novales, 1990), global functional expansions in
polynomials (Marcet, 1992; Judd, 1992), piecewise linear interpolation methods (Coleman, 1989; Baxter,
1991) and quadrature techniques (Tauchen and Hussey, 1991).
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
336
CALIBRATION
S126
F.CANOVA
Here we assume that either f is available analytically or that one of the existing numerical
procedures has been employed to obtain a functional which approximates f in some sense,
where γ=i(β) and
is a given norm. Given the model f,
i.e.
an approximation procedure a set of parameters β, and a probability distribution for Zt
(denoted by κ(Zt)), we can infer the model-based probability distribution for Xt.
be the density of the Xt vector, conditional on the parameters β and the
Let
be the density of the parameters, conditional on the information set
model f, let
available to the simulator and the model f, and let
be the joint density of
represents the probability that a particular
simulated data and of parameters.
path for the endogenous variables will be drawn given a parametric structure for the
summarizes the information on
artificial economy and a set of parameters, while
the parameters available to a researcher. Note that is assumed to be independent of and
π may depend on f, i.e. if we are using a GE model we may want to use only estimates
obtained with similar GE models. For a given β, Xt is random because Zt is random, i.e.
is a deterministic transformation of k(Zt).
Throughout this paper we are interested in studying the behaviour of functions of
simulated data (denoted by µ(X t )) under the predictive density
i.e. evaluating objects of the form:
(1)
where
and is the parameter space and is the support of the exogenous variables.
be the corresponding vector of functions of the actual data.
Let
The problem of measuring the fit of the model can be summarized as follows. How likely
? To answer note that from equation (1) we can compute
is the model to generate
probabilities of the form P(v(Xt)⑀D), where D is a bounded set and v(Xt) includes moments
and other statistics of the simulated data. To do this let µ(Xt)=␹(Xt, [Xt:v(Xt) ⑀D]) where ␹(Xt,
S) is the indicator function, i.e. ␹(Xt, S)=1 if v(Xt)⑀S and zero otherwise. Similarly, we can
construct quantiles q(Xt) by appropriately choosing D (see e.g. Geweke, 1989). Finally, we
satisfying
for any given vector v, by appropriately
can also find a
selecting the indicator function.
Model evaluation then consists of several types of statements which are interchangeable
and differ only in the criteria used to measure distance. First, we can compute
In other words, we can examine the likelihood of an event (the observed
realization of the summary statistics in the actual data) from the point of view of the model.
Extreme values for this probability indicate a poor ‘fit’ in the dimensions examined.
we can then choose the set
Alternatively, if we can measure the sampling variability of
plus one or two standard deviations and either
D to include the actual realization of
check whether is in D or calculate P[v(Xt)⑀D].
Implementation
There are four technical implementation issues which deserve some discussion. The first
concerns the computation of integrals like those in equation (1). When the (β, Zt) vector is
of high-dimension simple discrete grid approximations, spherical or quadrature rules
quickly become infeasible since the number of function evaluations increases exponentially
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
STATISTICAL INFERENCE IN CALIBRATED MODELS
STATISTICAL INFERENCE IN CALIBRATED MODELS
337
S127
with the dimension of β and Z t . In addition, unless the contours of
are of ellipsoidal forms, grid evaluations may explore
this density inefficiently. There are several feasible alternatives: one is the Monte Carlo
procedure described in Geweke (1989), another is the data-augmentation procedure of
Tanner and Wong (1987), a third is the ‘Gibbs sampler’ discussed in Gelfand and Smith
(1990). Finally, we could use one of the quasi-random procedures proposed by Niederreiter
(1988).
In this paper we adopt a Monte Carlo approach. After drawing with replacement i.i.d. β
vectors and Zt paths, we substitute sums over realizations for the integrals appearing in
equation (1) and appeal to the strong law of large numbers for functions of i.i.d. random
variables to obtain
(2)
where N is the number of replications. Note that, although
(and p) are, in general,
unknown, sampling from them can be conveniently accomplished by simulating the model
repeatedly for random (Zt, β), i.e. randomly drawing exogenous forces and selecting a
parameter vector and using the decision rule to compute time paths for Xt.
Second, since in most cases the function f is unknown, itself becomes unknown and
the direct computation of equation (1) is infeasible. If the approximation to f is accurate,
we could simply neglect the error and proceed using
in place of
where
is the joint density of simulated data and parameters using the
information set and the approximation rule However, since only little is known about
the properties of approximation procedures and some have only local validity (see e.g.
Christiano, 1990; Dotsey and Mao, 1991), we may want to explicitly account for the
existence of an approximation error in conducting inference. In this case, following Geweke
(1989), we would replace equation (1) with:
(3)
are weights which depend on the ‘true’ density
and on the
where
For example, if a quadratic approximation around
approximation density
the steady state is used, the density can be chosen so that draws of Zt inducing paths of
Xt which are in the tails of (i.e. paths which are very far away from steady states) receive
a very small weight in the calculation of the statistics of interest.
Third, we must specify a density for the parameters of the model. We could select it on
to be the asymptotic distribution of
the basis of one specific data set and specify
a GMM estimator (as in Burnside et al., 1993), of a simulated method of moments (SMM)
estimator (as in Canova and Marrinan, 1993), or of a ML estimator of β (as in Phillips,
1991). The disadvantage of these approaches is that the resulting density measures the
uncertainty surrounding β present in a particular data set and does not necessarily reflect
the uncertainty faced by a researcher in choosing the parameters of the model. As Larry
Christiano has pointed out to the author, once a researcher chooses the moments to match,
the uncertainty surrounding estimates of β is small. The true uncertainty lies in the choice
of moments to be matched and in the sources of data to be used to compute estimates.
so as to summarize efficiently all existing
A better approach would be to select
information, which may include point estimates of β obtained from different estimation
techniques, data sets, or model specifications. El-Gamal (1993a, b) has formally solved the
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
338
CALIBRATION
S128
F.CANOVA
problem of finding such a
using Bayesian methods. The resulting
is
the least informative pseudo-posterior density on the parameter space which is consistent
with a set of constraints describing the information contained in various estimation experiments. El-Gamal suggests a Gibbs sampler algorithm to compute this density but, in practice, there are simpler ways to construct empirical approximations to this type of density.
One would be to count estimates of β previously obtained in the literature and construct
by smoothing the resulting histogram. For example, if one of the elements of the
β vector is the risk aversion parameter, counting estimates obtained over the last 15 years
from fully specified general equilibrium models and smoothing the resulting histogram, we
would obtain a truncated (below-zero) bell-shaped density, centred around two and very
small massabove four. Alternatively, we could take what the profession regards as a reasonable range for β and assume more or less informative densities on the support depending on
available estimates. If theoretical arguments suggest that the maximum range for e.g. the
risk aversion parameter is [0, 20], we can put higher weights on the interval [1, 3] where
most of the estimates lie. If for some parameters previous econometric evidence is scant,
measurement is difficult, or there are no reasons to expect that one value is more likely than
others, we could assume uniform densities on the chosen support.
Available estimates of β are not necessarily independent (the same data set is used in
many cases) and some are less reliable than others. Non-independent estimates are
legitimate candidates to enter the information set as long as they reflect sampling variability
or different estimation techniques. The influence of less reliable estimates or of estimates
obtained with different model specifications can be discounted by giving them a smaller
weight in constructing histograms (see also El-Gamal, 1993a, b).
Finally, in many applications the joint density of the parameter vector can be factored
into the product of lower-dimensional densities. If no relationship across estimates of the
is the product of univariate densities. If estimates of
various parameters exists,
certain parameters are related (e.g. in the case of the discount factor and the risk aversion
parameter in asset pricing models), we can choose multivariate densities for these
dimensions and maintain univariate specifications for the densities of the other parameters.
To summarize, to implement the procedure we need to do the following:
•
•
•
•
•
where represents the information
Select a reasonable (data-based) density
set available to a researcher, and a density κ(Zt) for the exogenous processes.
and zt from κ(Zt).
Draw vectors β from
and compute µ(xt) using the model xt=f(zt, β)
For each draw of β and zt, generate
or the approximation
Repeat the two previous steps N times.
Construct the frequency distribution of µ(xt), compute probabilities, quantiles and
other measures of interest.
An Interpretation
The proposed framework of analysis lends itself to a simple Bayesian interpretation. In
this case
we treat as the prior on the parameters. The function is entirely
analogous to a classical likelihood function for Xt in a standard regression model. The
difference is that need not be the correct likelihood for and need not have a closed
form. Then equation (1) is the conditional expectation of µ(Xt) under the predictive
density of the model and probability statements based on equation (1) can be justified
using the arguments contained in Box (1980).
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
STATISTICAL INFERENCE IN CALIBRATED MODELS
STATISTICAL INFERENCE IN CALIBRATED MODELS
339
S129
There is also a less orthodox interpretation of the approach which exchanges the role of
and
and is nevertheless reasonable. In this case
is the
prior and represents the a priori degree of confidence posed by the researcher on the time
summarizes the
path generated by the model given the parameters while
information contained in the data. Then equation (1) is a ‘pseudo-posterior’ statement
about the model’s validity once the empirical evidence on the parameters is taken into
account.
It is useful to note that, if we follow the first approach, we can relate the proposed
to the data-based priors employed in Meta-Analysis (see Wolf,
construction of
1986) and in the ‘consensus literature’ (see e.g. Genest and Zideck, 1986). El-Gamal (1993a)
spells out in detail the connection with these two strands of literature.
3. ROBUSTNESS ANALYSIS
If we adopt a Monte Carlo approach to compute simulated densities for the statistics of
interest, an automatic and efficient global sensitivity analysis is performed on the support of
the parameter space as a by-product of the simulations. Sensitivity analysis, however, can
take other more specific forms. For example, we may be interested in examining how likely
when β is fixed at some prespecified value This would be the
µ(Xt) is to be close to
case, for example, if β includes parameters which can be controlled by the government and
is e.g. the current account balance of that country. In this case we could choose a path
for Zt and analyse the conditional distribution of µ(Xt) for the selected value(s) of β. Alternatively, we might wish to assess the maximal variation in µ(Xt) consistent, say, with β being
within two standard deviations of a particular value. Here we choose a path for Zt and
construct paths for µ(Xt) for draws of β in the range. Finally, we may be interested in
knowing which dimensions of β are responsible for particular features of the distribution of
µ(Xt). For example, if the simulated distribution of µ(Xt) has a large spread or fat tails, a
researcher may be interested in knowing whether technology or preference parameters are
responsible for this feature. In this case we would partition β into [β1, β2] and compute the
where
is a prespecified value (or
simulated distribution of µ(Xt) conditional on
set of values.
So far, we have examined the robustness of the results to variations of the parameters
within their support. In some cases it is necessary to study the sensitivity of the results to
local perturbations of the parameters. For example, we may be interested in determining
how robust the simulation results are to changes of the parameters in a small
neighbourhood of a particular vector of calibrated parameters. To undertake this type of
analysis we can take a numerical version of the average derivative of µ(X t) in the
neighbourhood of a calibrated vector (see Pagan and Ullah, 1991). Because global and local
analyses aim at examining the sensitivity of the outcomes to perturbations in the parameters
of different size they provide complementary information and should both be used as
specification diagnostics for models whose parameters are calibrated.
4. A COMPARISON WITH EXISTING METHODOLOGIES
The framework of analysis in Section 2 is general enough to include simulation undertaken
after the parameters are both calibrated and estimated via method of moments as special
is a deterministic transforcases. To show this it is convenient to recall that
The two procedures can then be recovered by
mation of
and, in some cases, also on
imposing restrictions on the shape and the location of
the shape and the location of k(Zt).
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
340
S130
CALIBRATION
F.CANOVA
Calibration exercises impose a point mass for
on a particular value of β (say,
). One interpretation of this density selection is that a simulator is perfectly confident in
the vector β used and does not worry about the cross-study or time-series uncertainty
surrounding estimates of β. In certain situations a path for the vector of exogenous
variables is also selected in advance either by drawing only one realization from their
distribution or by choosing a zt on the basis of extraneous information, e.g. inputting
Solow’s residuals in the model, so that κ(Zt) is also a singleton. In this instance, the density
of µ(Xt) has a point mass and because the likelihood of the model to produce any event is
either 0 or 1, we must resort to informal techniques to assess the discrepancy of simulated
and actual data. In some studies the randomness in Zt is explicitly considered, repeated
draws for the exogenous variables are made for a fixed and moments of the statistics of
interest are computed by averaging the results over a number of simulations (see e.g.
Backus et al., 1989).
Simulation exercises undertaken with estimation of the parameters are also special cases
has a point mass at β*, where β* is either the GMM
of the above framework. Here
estimator, the SMM estimator (see Lee and Ingram, 1991), or the simulated quasimaximum likelihood (SQML) estimator of β (see Smith, 1993). Simulations are typically
and standard errors for µ(Xt)
performed by drawing one realization from
are computed using the asymptotic standard errors of β* and the functional form for µ. In
some cases,
is taken to be the asymptotic distribution of one of the above estimators
(e.g. Canova and Marrinan, 1993). In this case, simulations are performed by drawing from
and the distance of simulated and actual data is computed using
measures of discrepancy like the ones proposed here.
In assessing the model’s performance this last set of procedures has two advantages over
calibration. First, they allow formal statements on the likelihood of selected parameter
values to reproduce the features of interest. For example, if a four standard deviations range
around the point estimate of the AR(1) parameter for the productivity disturbance is [0·84,
0·92], then it is highly unlikely (with a probability higher than 99%) that a unit root
productivity disturbance is needed to match the data. Second, they provide a set-up where
sensitivity analysis can easily be undertaken (although not often performed).
These procedures, however, have also two major shortcomings. First, they impose a
strong form of ignorance which does not reflect available a priori information. The vector β
may include meaningful economic parameters which can be bounded on the basis of
theoretical arguments but the range of possible β with GMM, SMM, or SQML procedures
is [-∞, ∞]. By appropriately selecting a hypercube for their densities a researcher can make
‘unreasonable’ parameter values unlikely and avoid a posteriori adjustments. Second,
simulations conducted after parameters are estimated may not constitute an independent
way to validate the model because the parameter estimates are obtained from the same data
set which is used later to compare results.
Simulation procedures where parameters are selected using a mixture of calibration and
estimation strategies have recently been employed by e.g. Heaton (1993) and Burnside et al.
(1993). Here some parameters are fixed using extraneous information while others are
formally estimated using moment (or simulated moment) conditions. Although these
strategies allow a more formal evaluation of the properties of the model than pure
calibration procedures, they face two problems. First, as in the case when the parameters are
all selected using GMM, SMM, and SQML procedures, the evaluation of the model is
problematic because measures of dispersion for the statistics of interest are based on one
data set and do not reflect the uncertainty faced by a simulator in choosing the unknown
features of the model. Second, as Gregory and Smith (1989) have pointed out, the smallsample properties of estimators obtained from these strategies may be far from reasonable
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
STATISTICAL INFERENCE IN CALIBRATED MODELS
STATISTICAL INFERENCE IN CALIBRATED MODELS
341
S131
unless calibrated parameters consistently estimate the true parameters. When this
condition is not satisfied, estimates of the remaining parameters are sensitive to errors in
pre-setting and results are misleading.
The Monte Carlo methodology we employ to evaluate the properties of the model is
related to those of Kwan (1990) and Gregory and Smith (1991) but several differences need
to be emphasized. First, Gregory and Smith take the model as a testable null hypothesis
while this is not the case here. Second, they do not account for parameter uncertainty in
evaluating the outcomes of the model. Third, because they take a calibrated version of the
model as the ‘truth’, they conduct sensitivity analysis inefficiently, by replicating
experiments for different calibrated values of the parameters. Kwan (1990) allows for
parameter uncertainty in his simulation scheme, but, following an orthodox Bayesian
approach, he chooses a subjective prior density for the parameters. In addition, he evaluates
the outcomes of the model in relative terms by comparing two alternative specifications
using a posterior-odds ratio: a model is preferred to another if it maximizes the probability
that the simulated statistics are in a prespecified set.
The procedure for sensitivity analysis we proposed extends the approach that Harrison
and Vinod (1989) used in deterministic computable general equilibrium models. The major
difference is that in a stochastic framework parameter uncertainty is only a part of the
randomness entering the model and the uncertainty characterizing the exogenous
processes is important in determining the randomness of the outcomes.
To conclude, we should mention that, parallel to the literature employing Monte Carlo
methods to evaluate calibrated models, there is also a branch of the literature which uses
alternative tools to examine the fit of calibrated models. This is the case e.g. of Smith (1993),
Watson (1993), and Canova et al. (1993) which assess the relevance of theoretical models
with regression R2’s, tests based on restricted and unrestricted VARs, and encompassing
procedures.
5. TWO EXAMPLES
The Equity Premium Puzzle
Mehra and Prescott (1985) suggest that an asset-pricing model featuring complete markets
and pure exchange cannot simultaneously account for the average risk-free rate and the
average equity premium experienced by the US economy over the sample 1889–1978 with
reasonable values of the risk aversion parameter and of the discount factor.
The model they consider is a frictionless Arrow-Debreu economy with a single
representative agent, one perishable consumption good produced by a single productive
unit or a ‘tree’, and two assets, an equity share and a risk-free asset. The tree yields a random
dividend each period and the equity share entitles its owner to that dividend in perpetuity.
The risk-free asset entitles its owner to one unit of the consumption good in the next period
only. The agent maximizes:
(4)
subject to:
(5)
where E0 is the mathematical expectation operator conditional on information at time zero,
␪ is the subjective discount factor, ␻ is the risk aversion parameter, ct is consumption, yt is the
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
342
S132
CALIBRATION
F.CANOVA
tree’s dividend, and are the prices of the equity and the risk-free asset, and et and ft are
the agent’s equity and risk-free asset holding at time t, Production evolves according to
yt+1=xt+1yt where xt, the gross growth rate, follows a two-state ergodic Markov chain with
Defining the states of the problem as (c, i) where yt=c
probability
and xt=λi, the period t equilibrium asset prices are
(6)
(7)
When the current state is (c, i), the expected equity return and the risk-free rate are:
(8)
(9)
The unconditional (average) expected returns on the two assets are
and the average equity premium is EP=Re-Rf, where πi are the Markov chain
and Σiπi=1, where
stationary probabilities, satisfying
Mehra and Prescott specified the two states for consumption (output) to be λ1=1+µ+v;
λ2=1+µ-v and restricted
and
They calibrated the three
technology parameters so that the mean, the standard deviation, and the AR(1) coefficient
of the model’s consumption match those of the growth rate of annual US consumption
over the period 1889–1978 and searched for combinations of the preference parameters (␪,
␻) in a prespecified interval to obtain values for the risk-free rate and the equity premium.
Given that the average, the standard deviations, and the AR(1) coefficient of annual growth
rate of US consumption are 1·018, 0·036, and -0·14, the implied values of µ, v, and ␾ are
0·018, 0·036, and 0·43, respectively. The range for ␻ was selected to be [0, 10] and this choice
was justified citing a number of empirical estimates of this parameter (see Mehra and
Prescott, 1985, p. 154). The range for ω was chosen to be [0, 1], but simulations which
produced a risk-free rate in excess of 4% were eliminated on the grounds that 4%
constitutes an upper bound consistent with historical experience. The puzzle is that the
largest equity premium generated by the model is 0·35%, which occurred in conjunction
with a real risk-free rate of about 4%, while the US economy for the period 1889–1978
experienced an annual average equity premium of 6·18% and an average real risk-free rate of
0·80%.
Two hidden assumptions underlie Mehra and Prescott’s procedure. First, they believe
that technology parameters can be tightly estimated while the uncertainty surrounding the
choice of preference parameters is substantial. Consequently, while the sensitivity of the
results is explored to variations in θ and ω within the range, no robustness check is made
for perturbations of the technology parameters. Second, by providing only the largest value
generated, they believe that it is a sufficient statistic to characterize the puzzle.
Here we repeat their exercise with three tasks in mind: first, to study whether the
uncertainty present in the selection of the technology parameters is important in
determining the magnitude of the puzzle; second, to formally measure the discrepancy of
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
STATISTICAL INFERENCE IN CALIBRATED MODELS
STATISTICAL INFERENCE IN CALIBRATED MODELS
343
S133
the model from the data using a variety of statistics based on the probability distribution of
outcomes of the model; third, to evaluate the contribution of two alleged solutions to the
equity premium puzzle proposed in the literature.
This first example is particularly simple since we have an exact solution for the
endogenous variables of the model. In addition, because the model produces values for the
mean of Rf and EP, variations in Xt are entirely determined by variations in β, so that
is proportional to
Therefore, once
we have selected
we can immediately determine the distribution of simulated means of the Rf-EP pair.
To select the density for the five parameters of the model we proceed in two steps. First,
we choose a maximum range for the support of β on the basis of theoretical considerations.
Table I. Equity premium puzzle
Note:
Pr 1 refers to the frequency of simulations for which the pair (Rf, EP) is in a classical 95% region around the actual values. Pr
2 reports the percentile of the simulated distribution where the actual (Rf, EP) pair lies. Pr 3 reports the probability that the
model generates values in each of the four quadrants delimited by the actual pair of (Rf, EP). Q1 is the region where Rf1>Rf
and EPs<EP, Q2 is the region where Rfs>Rf and EP4ⱖEP, Q3 is the region where RfsⱕRf and EP4<EP and Q4 is the region where
RfsⱕRf and EPsⱖEP.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
344
S134
CALIBRATION
F.CANOVA
Second, we specify the joint density to be the product of five univariate densities and select
each univariate density to be a smoothed version of the frequency distribution of estimates
existing in the literature. The densities and their support are in panel A of Table I. Therange
for ω is the same as that of Mehra and Prescott and the chosen χ2 density has a mode at 2,
where most of the estimates of this parameter lie, and a low mass (smaller than 5%) for
values exceeding 6. The range for θ reflects the results of several estimation studies which
obtained values for the steady-state real interest rate in the range [-0·005, 0·04] (see e.g.
Altug, 1989; Dunn and Singleton, 1986; or Hansen and Singleton, 1983) and of simulation
exercises which have a steady-state real interest rate in the range [0, 0·05] (see e.g. Kandel
and Stambaugh, 1990; or Mehra and Prescott, 1985). The density for ␪ is skewed to express
the idea that a steady-state real interest rate of 2–3% or lower is more likely than a steadystate interest rate in excess of 4%. Note that although we assume that the densities of θ and
ω are independent, many estimates of these two parameters are not. However, the rank
correlation coefficient for the pairs of estimates is small and none of the results we present
depends on this simplifying assumption.
To provide a density for µ, v and ␾ we experimented with two procedures. The first,
which is used in the basic experiment, involves taking the 10 sub-sample estimates of the
mean, of the standard deviation, and of the AR(1) coefficient of the growth rate of
consumption over 10-year samples contained in Mehra and Prescott (1985, p. 147) as
characterizing reasonable consumption processes and then constructing a uniform discrete
density over these triplets. The second involves dividing the growth rates of consumption
over the 89 years of the sample into two states (above and below the mean), estimating a
measure of dispersion for the first two moments and for the AR(1) coefficient of the growth
rate of consumption in each state and directly inputting these estimates into the model. In
this case simulations are performed by assuming a joint normal density for the mean, the
standard deviation, and AR(1) coefficient in each state centred around the point estimate of
the parameters and maximum support within two standard deviations of the estimate.
Figures 1–4 present scatterplots of the simulated pairs (Rf, EP) when 10,000 simulations
are performed. We summarize the features of the joint distribution in panel B of Table I
Figure 1. Scatterplot risk-free rate-equity premium: Mehra-Prescott case
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
STATISTICAL INFERENCE IN CALIBRATED MODELS
STATISTICAL INFERENCE IN CALIBRATED MODELS
Figure 2. Scatterplot risk-free rate-equity premium: basic case
Figure 3. Scatterplot risk-free rate-equity premium: beta>1 case
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
345
S135
346
S136
CALIBRATION
F.CANOVA
Figure 4. Scatterplot risk-free rate-equity premium: Reitz case
using a number of statistics. To evaluate the discrepancy of the model from the data we
report (1) the probability that the model generates values for (Rf, EP) which fall within a two
standard deviation band of the actual mean, (2) the percentile contour of the simulated
distribution where the actual means of (Rf, EP) lies, and (3) the probability that the
simulated pair is in each of the four quadrants of the space delimited by the actual means of
(Rf, EP).
Figure 1 reports the scatterplot obtained with the Mehra and Prescott specification (i.e.
when technology parameters are fixed and we draw replications from the densities of θ and
ω only). It is necessary to check that the maximum value of the equity premium consistent
with a risk free-rate not exceeding 4% is only 0·0030, confirming Mehra and Prescott’s
conclusion. Also for this specification, the distribution of the model’s outcomes is uniform
and the mode of the joint distribution (the most likely value from the point of view of the
model) is at Rf=0·110, EP=0·0094. The probabilistic measures of discrepancy suggest that
a large portion of the simulations are in the region where the simulated Rf exceeds the mean
of Rf and the simulated EP is below the mean of EP we find in the data, that about 73% of
the simulations produce pairs within a classical 95% ball around the actual means of (Rf,
EP), and that the actual mean pair is outside the 99 percentile contour.
Figure 2 reports the scatterplot obtained with the basic specification of the model. Also
in this case, the puzzle, as defined by Mehra and Prescott, is evident: if we set a 4% upper
bound to the risk-free rate, the maximum equity premium generated is only 0·0038.
However, with this specification, the distribution is bimodal and most of the simulated
pairs lie on a ridge parallel to the Rf axis. The probability that the model generates values in
a ball centred around the actual means of (Rf, EP) is now 81·4%. However, in 100% of the
cases the simulated risk-free rate exceeds the actual mean and the simulated equity
premium is below the actual mean and the actual pair still lies outside the 99 percentile
contour of simulated distribution.
To examine whether the selection of the density for the technology parameters has
effects on the results, we also conducted simulations using the alternative distribution for
these parameters. No substantial changes emerge. For example, the probability that the
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
STATISTICAL INFERENCE IN CALIBRATED MODELS
STATISTICAL INFERENCE IN CALIBRATED MODELS
347
S137
model generates pairs in a ball centred around the actual means of (Rf, EP) is 80·3% and the
maximum value for EP compatible with a Rf not exceeding 4% is 0·0025.
Several conclusions can be drawn from this first set of exercises. First, even after taking
into account the uncertainty surrounding estimates of the technology parameters, the
puzzle remains regardless of the way it is defined (maximum values, modes, or contour
probabilities): the model cannot generate (Rf, EP) pairs which match what we see in the
data. Second, once the uncertainty surrounding estimates of the technology parameters is
taken into account, the simulated distributions are bimodal, highly left skewed, and have a
fat left tail, indicating that lower than average values are more probable and that very small
values have nonnegligible probability. Third, the simulated risk-free rate is always in excess
of the actual one, a result that Weil (1990) has termed the risk-free rate puzzle. Fourth, while
the model fails to generate values for (Rf, EP) which replicate the historical experience, in
more than 80% of the simulations it produces pairs which are within two standard
deviations of the actual means.
Next, we conduct two exercises designed to examine the contribution of the
modifications suggested by Kocherlakota (1990), Benninga and Protopapadakis (1990), and
Rietz (1988) to the solution of the puzzle. The first experiment allows the discount factor θ
to take on values greater than 1. The justification is that, in a growing economy, reasonable
values for the steady-state real interest rate can be obtained even with θ greater than 1. In
this experiment we still maintain the truncated normal density for θ used in the baseline
case but increase the upper value for its range to 1·04 and allow about 10% of the density in
the region above 1.0.
The second experiment assumes the presence of a third unlikely crash state where
consumption falls substantially.2 The justification for including a third state is that in the
Great Depression consumption fell substantially and excluding such a state may have
important implications on the results (a conclusion denied by Mehra and Prescott, 1988).
With this specification there are two new parameters which cannot be measured from
available data: ξ, the probability of a crash state and ξ, the percentage fall in consumption in
the crash state. Rietz (1988) searched over the a priori ranges of [0·0001, 0·2] and [µ/(1+µ), 1v/(1+µ)] and examined the magnitude of the maximum simulated equity premium that the
model consistent with a simulated risk-free rate below 4%. We maintain these ranges in our
experiment and assume on these supports an exponential density for ξ and a three-point
discrete density for ξ summarizing the three cases examined by Rietz.
Allowing the discount factor to take on values greater than 1 goes a long way towards
reducing the discrepancy of the model from the data (see Figure 3) since it shifts the
univariate distribution of Rf towards negative values (the minimum and maximum values
of Rf are now (-0·084, 0·0.092). For example, the probability that the model generates pairs in
a ball centred around the actual means of (Rf, EP) is now 85·7% and in only 7·4% of the cases
is the simulated risk-free rate in excess of the actual means. Because of this shift in the
univariate distribution of Rf, the maximum value of EP consistent with a risk-free rate
below 4% is now 0·031. Despite these differences, the location and the shape of the
univariate distribution of EP are unaffected. Hence, although the equity premium puzzle is
‘solved’ when defined in terms of the maximum simulated EP consistent with a simulated
Rf below 4%, it is still very evident when we look at the distributional properties of the
simulated EP.
2
The three consumption states are
and the transition matrix
has elements:
Note
that the experiment is conceptually different from the previous ones since there are two extra degrees of freedom
(the new parameters ξand ξ) and no extra moments to be matched.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
348
CALIBRATION
S138
F.CANOVA
The second modification is much less successful (see Figure 4). It does shift the
univariate distribution of EP to the right (the mode of 0·035) and increases the dispersion
of simulated EPs but it achieves this at the cost of shifting the distribution of Rf towards
unrealistic negative value (the mean is -0·15 and the 90% range is [-0·940, 0·068]) and of
scattering the simulated (Rf, EP) pairs all over the place. For example, the probability that the
simulated pair is in a ball centred around the actual means of (Rf, EP) decreases to 72·7%
and the probabilities that the model generates values in each of thefour quadrants delimited
by the actual means of (Rf, EP) are almost identical. Finally, the maximum EP consistent
with a Rf below 4% is 0·747. Therefore, adding a crash state shifts the mode and stretches
and tilts the shape of the joint simulated distribution. Roughly speaking, too many (Rf, EP)
configurations now have equal probability, and this weakens the ability of the theory to
provide a coherent answer to the question posed.
Technology Shocks and Cyclical Fluctuations in GNP
Kydland and Prescott (1982) showed that a one-sector growth model driven by technology
shocks calibrated to reproduce the statistical properties of Solow residuals explains about
70% of the variance of per capita US output. This result has spurred much of the subsequent literature which tries to account for business cycle regularities in models where
monetary impulses play no role (the so-called real business cycle literature). Kydland and
Prescott’s initial estimate has been refined by adding and subtracting features to the basic
model (see Hansen, 1985) but the message of their experiment remains: a model where
technology shocks are the only source of disturbance explains a large portion of the variability of per capita US output.
Recently, Eichenbaum (1991) has questioned this assertion because ‘decisions based
and
and
solely on the point estimate of λy are whimsical
var(yt) are the variance of the cyclical component of simulated and actual output) and
suggests that ‘the model and the data, taken together, are almost completely uninformative
about the role of technology shocks in generating fluctuations in US output’ (pp. 614–615).
Using an exactly identified GMM procedure to estimate the free parameters, he finds that
the model explains anywhere between 5% and 200% of the variance of per capita US output.
In this section we repeated Eichenbaum’s exercise with three goals in mind. First, we are
interested in knowing that is the most likely value of λy from the point of view of the model
(i.e. in locating the mode of the simulated distribution). Second, we want to provide
confidence bands fo λy which reflect the uncertainty faced by a researcher in choosing the
parameters of the model (not the uncertainty present in the data, as in Eichenbaum). Third,
we wish to verify whether normal confidence bands appropriately describes the uncertainty
surrounding point estimates of λy and examine which feature of the model make deviations
from normality more evident.
The model is the same as Eichenbaum’s and is a simple variation of Hansen’s (1985)
model which allows for deterministic growth via labour-augmenting technological
progress. The social planner of this economy maximizes
(10)
subject to:
(11)
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
STATISTICAL INFERENCE IN CALIBRATED MODELS
STATISTICAL INFERENCE IN CALIBRATED MODELS
349
S139
where ct is per capital consumption, T-ht is leisure, and kt the capital stock. When δ is
different from 1, a closed-form stationary solution to the problem does not exist. Here we
compute an approximate decision rule for the endogenous variables using a loglinear expansion around the steady state after variables have been linearly detrended as in King et al.
(1988), but we neglect the approximation error in constructing probability statements on
and no weighting).
the outcomes of the model (i.e. we use
There are seven parameters in the model, five deep (δ, the depreciation rate of capital; β,
the subjective discount rate; ␺, leisure’s weight in the utility function; α, labour’s share in
output; γ, the constant unconditional growth rate of technology) and two which appear
only because of the auxiliary assumptions we made on the stochastic process for technology
shocks (ρ, the AR parameter and σ the standard deviation of the shock). Hansen (1985)
calibrated these seven parameters (the values are in the first column of panel A of Table II)
and found that λy≈1. Eichenbaum (1991) estimated all parameters except β (which is
calibrated) using a method of moments estimator (estimates and standard deviations are in
the second column of panel A of Table II) and found (1) a point estimate of λy of 0·80, (2) a
large standard deviation about the point estimate of λy due primarily to the uncertainty
surrounding estimates of ρ and σ, and (3) a strong sensitivity of the point estimate of λy to
small perturbations in the parameter vector used.
Table II. Technology shocks and cyclical fluctuations in GNP
Note:
Estimated standard errors are in parentheses. Pr 1 refers to the frequency of simulations for which the variance of simulated
output is in a classical 95% region around the actual value of the variance of detrended output. Pr 2 reports the percentile of
the simulated distribution where the point estimate of the actual variance of output lies.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
350
S140
CALIBRATION
F.CANOVA
In the exercise we conduct, we assume that
is the product of seven univariate
densities. Their specification appear in the third column of panel A of Table II The range for
the quarterly discount factor corresponds to the one implied by the annual range used in
the previous example and the density is the same. ␦ is chosen so that the annual depreciation
rate of the capital stock is uniformly distributed between 8% and 12% per year. The range
is selected because in simulation studies ␦ is commonly set to 0·025, which corresponds to
a 10% annual depreciation rate, while estimates of this parameter lie around this value (e.g.
McGratten et al., 1991, have a quarterly value of 0·0310 and a standard deviation of 0·0046,
while Burnside et al., 1993, have a quarterly value of 0·0209 and a standard deviation of
0·0003). The range for a reflects calculations appearing in Christiano (1988) where,
depending on how proprietors’ income is treated, the share of total output paid to capital
varies between 0·25 and 0·43, and the estimate obtained, among others, in McGratten et al.
(1991). We chose the densities for ρ and σ as in Eichenbaum because the econometric
evidence on these two parameters is scant andthe values used in most simulation studies
fall within a one standard deviation band around the mean of the assumed density (see e.g.
Kydland and Prescott, 1982; Hansen, 1985). Finally, T is fixed at 1369 hours per quarter,
the density for γ matches the quarterly distribution of unconditional quarterly growth
rates of US output for the period 1950–1990, and ␺ is endogenously chosen so that the
representative household spends between one sixth and one third of its time working in the
steady state.
We performed 1000 simulations with time series of length T=124 and filtered both
simulated and actual GNP data with the Hodrick and Prescott filter.3 The results appear in
panel B of Table II and in Figure 5, where we present a smoothed version of the simulated
distribution of λy. The distribution is scaled so that with the point estimates of the
parameters used by Eichenbaum λy=0·80. The implied value of λy using Hansen’s
parameters is 0·84.
Figure 5. Density of variance ratio: HP filtered data
3
We use the Hodrick and Prescott filter to maintain comparability with previous work. The results obtained
when the data are linearly detrended or first-order differenced are not substantially different.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
STATISTICAL INFERENCE IN CALIBRATED MODELS
STATISTICAL INFERENCE IN CALIBRATED MODELS
351
S141
The mode of the distribution of λy is at 0·9046, the mean at 0·8775, and the median at
0·5949. The dispersion around these measures of location is very large. For example, the
standard deviation is 0·7635 and the 90% range of the distribution is [0·2261, 2·6018]. The
simulated distribution is far from normal and its right tail tends to be very long. Hence the
range of reasonable values of λy is very large, and, as in Eichenbaum, small perturbations in
the parameter vector induce large variations in the variance ratio. In addition, normal
confidence bands do not appropriately characterize the uncertainty surrounding the
outcomes of the model.
Several other features of the simulated distribution are worth mentioning. First, in
67·3% of the cases the variance of simulated output is smaller than the variance of actual
output. Second, in 42·7% of the simulations the variance of simulated output is within a
95% confidence interval centred around the estimate of the variance of actual output.
i.e. is the median
Third, if we select v=0·5 and look for the satisfying
of the simulated distribution, we find that the median value of the variance of simulated
GNP is outside the 95% normal confidence interval for the variance of actual GNP.
When we ask which parameter is responsible for the wide dispersion in the estimates of
λy, we find that it is the location and width of the support of ρ which induce this feature in
the distribution of λy. For example, assuming that the density of ρ has a point mass at 0·94
and maintaining the same densities for the other parameters, we find that location
measures of the simulated distribution of λy decrease (the mode is now at 0·792) and the
standard deviation drops to 0·529. Similar conclusions are obtained by shifting the range of
ρ towards 0·90 or by cutting the range of possible ρ in half without changing the mean
value. Hence, as in Eichenbaum, we find that it is the uncertainty present in the choice of
the parameters of the exogenous processes rather than the uncertainty present in the
selection of the deep parameters of the model that is responsible for the large spread in the
distribution of λy.
6. CONCLUSIONS
This paper describes a Monte Carlo procedure to evaluate the properties of calibrated
general equilibrium models. The procedure formalizes the choice of the parameters and the
evaluation of the properties of the model while maintaining the basic approach used in
calibration exercises. It also realistically accounts for the uncertainty faced by a simulator in
choosing the parameters of the model. The methodology allows for global sensitivity analysis for parameters chosen within the range of existing estimates and evaluates the discrepancy of the model from the data by attaching probabilities to events a simulator is interested
in characterizing. The approach is easy to implement and includes calibration and simulation exercises conducted after the parameters are estimated by simulation and GMM techniques as special cases. We illustrate the usefulness of the approach as a tool to evaluate the
performance of theoretical models with two examples which have received much attention
in the recent macroeconomic literature: the equity premium puzzle and the ability of a real
business cycle model to reproduce the variance of actual US output. Finally, it is worth
noting that for problems of moderate size, the computational complexity of the procedure is
limited. For both examples presented the entire Monte Carlo routine required about a
minute on a 486–33 MHz machine using RATS386 programs.
ACKNOWLEDGEMENTS
Part of this research was undertaken while the author was also associated with the European University Institute, Florence. The author has benefited from the comments and the
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
352
S142
CALIBRATION
F.CANOVA
suggestions of a large number of colleagues, including two anonymous referees, David
Backus, Larry Christiano, Frank Diebold, Javier Diaz, Mahmoud El-Gamal, John Geweke,
Eric Ghysels, Bruce E.Hansen, Gary Hansen, Jane Marrinan, Yaw Nyarko, Adrian Pagan,
Franco Peracchi, Victor Rios, Gregor Smith, Herman Van Dijk, and Randall Wright. He
would also like to thank the participants of seminars at Brown, European University Institute, NYU, Montreal, Rochester, Penn, University of Rome, Carlos III Madrid, Free University Bruxelles, CERGE, University of Minnesota, University of Maryland, Summer
Meetings of the North American Econometric Society and the Conference on ‘Econometric
Inference Using Simulation Techniques’ held in Rotterdam on 5–6 June 1992 for useful
discussions.
REFERENCES
Altug, S. (1989), ?‘Time to build and aggregate fluctuations: some new evidence’, International
Economic Review, 30, 889–920.
Backus, D., A. Gregory and A.Zin (1989), ‘Risk premiums in the terms structure: evidence
from artificial economies’, Journal of Monetary Economics, 24, 371–399.
Baxter, M. (1991), ‘Approximating suboptimal dynamic equilibria: an Euler equation approach’,
Journal of Monetary Economics, 27, 173–200.
Benninga, S. and A. Protopapadakis (1990), ‘Leverage, time preference, and the equity
premium puzzle’, Journal of Monetary Economics, 25, 49–58.
Box, G. (1980), ‘Sampling and Bayes inference in scientific modelling and robustness’,
Journal of the Royal Statistical Society, Ser. A, 143, 383–430.
Burnside, C., M.Eichenbaum and S.Rebelo (1993), ‘Labor hoarding and the business cycle’,
Journal of Political Economy, 101, 245–273.
Canova, F. and J.Marrinan (1993), ‘Profits, risk and uncertainty in foreign exchange markets’,
Journal of Monetary Economics, 32, 259–286.
Canova, F., M. Finn and A.Pagan (1993), ‘Evaluating a real business cycle model’,
forthcoming, in C. Hargreaves (ed.), Nonstationary Time Series Analyses and Cointegration,
Oxford: Oxford University Press.
Christiano, L. (1988), ‘Why does inventory investment fluctuate so much?’ Journal of
Monetary Economics, 21, 247–280.
Christiano, L. (1990), ‘Solving the stochastic growth model by linear quadratic approximation
and by value function iteration’, Journal of Business and Economic Statistics, 8, 23–26.
Coleman, W. (1989), ‘An algorithm to solve dynamic models’, Board of Governors of the
Federal Reserve System, International Finance Division, Discussion Paper No. 351.
Dotsey, M. and C.S.Mao (1991), ‘How well do linear approximation methods work? Results
for suboptimal dynamic equilibria’, Journal of Monetary Economics, 29, 25–58.
Dunn, D. and K.Singleton (1986), ‘Modelling the term structure of interest rates under
non-separable utility and durability of goods’, Journal of Financial Economics, 17, 27–55.
Eichenbaum, M. (1991), ‘Real business cycle theory: wisdom or whimsy?’ Journal of Economic
Dynamic and Control, 15, 607–621.
El-Gamal, M. (1993a), ‘The extraction of information from multiple point estimates’,
forthcoming in Journal of Nonparametric Statistics.
El-Gamal, M. (1993b), ‘A Bayesian interpretation of extremum estimators’, manuscript,
California Institute of Technology.
Fair, R. (1991), ‘Estimating event probabilities from macroeconomic models using stochastic
simulations’, Yale University, manuscript.
Friedman, M. (1959), Essays in Positive Economics, New York: Aldine Press.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
STATISTICAL INFERENCE IN CALIBRATED MODELS
STATISTICAL INFERENCE IN CALIBRATED MODELS
353
S143
Frisch, R. (1933), ‘Propagation problems and impulse problems in dynamics economies’, in
Economic Essays in Honor of Gustav Cassel, London.
Geweke, J. (1989), ‘Bayesian inference in econometric models using Monte Carlo integration’,
Econometrica, 57, 1317–1339.
Gelfand, A. and A.Smith (1990), ‘Sampling based approaches to calculating marginal densities’,
Journal of the American Statistical Association, 85, 398–409.
Genest, C. and M.Zidak (1986), ‘Combining probability distributions: a critique and an
annotated bibliography’, Statistical Science, 1, 114–148.
Gregory, A. and G.Smith (1989), ‘Calibration as estimation’, Econometric Reviews, 9(1), 57–
89.
Gregory, A. and G.Smith (1991), ‘Calibration as testing: inference in simulated macro
models’, Journal of Business and Economic Statistics, 9(3), 293–303.
Haavelmo, G. (1944), ‘The probability approach in econometrics’, Econometrica, 12,
Supplement.
Hansen, G. (1985), ‘Indivisible labor and the business cycle’, Journal of Monetary Economics,
16, 309–328.
Hansen, L. and T.Sargent (1979), ‘Formulating and estimating dynamic linear rational
expectations models’, Journal of Economic Dynamic and Control, 2, 7–46.
Hansen, L. and K.Singleton (1983), ‘Stochastic consumption, risk aversion and temporal
behavior of asset returns’, Journal of Political Economy, 91, 249–265.
Harrison, G. and H.D.Vinod (1989), ‘Sensitivity analysis of applied general equilibrium
models: completely randomized factorial sampling designs’, University of New Mexico,
manuscript.
Heaton, J. (1993), ‘The interaction between time nonseparable preferences and time
aggregation’, Econometrica, 61, 353–381.
Journal of Business and Economic Statistics, January 1990.
Judd, K. (1992), ‘Projection methods for solving aggregate growth models’, Journal of Economic
Theory, 58, 410–452.
Kandel, S. and R.Stambaugh (1990), ‘Expectations and volatility of consumption and asset
returns’, Review of Financial Studies, 3, 207–232.
King, R., C.Plosser and S.Rebelo (1988), ‘Production, growth and business cycles: I and II’,
Journal of Monetary Economics, 21, 195–232 and 309–342.
Kocherlakota, N. (1990), ‘On the discount factor in growth economies’, Journal of Monetary
Economics, 25, 45–48.
Kydland, F. (1992), ‘On the econometrics of world business cycles’, European Economic
Review, 36, 476–482.
Kydland, F. and E.Prescott (1982), ‘Time to build and aggregate fluctuations’, Econometrica,
50, 1345–1370.
Kydland, F. and E.Prescott (1991), ‘The econometrics of the general equilibrium approach
to business cycles’, The Scandinavian Journal of Economics, 93(2), 161–178.
Kwan, Y.K. (1990), ‘Bayesian calibration with an application to a non-linear rational expectation
two country model’, mimeo, University of Chicago Business School.
Lee, B.S. and B.Ingram (1991), ‘Simulation estimators of time series models’, Journal of
Econometrics, 47(2/3), 197–206.
Marcet, A. (1992), ‘Solving nonlinear stochastic models by parametrizing expectations: an
application to asset pricing with production’, Universitat Pompeu Fabra, Working Paper 5.
Mehra, R. and E.Prescott (1985), ‘The equity premium: a puzzle’, Journal of Monetary Economics,
15, 145–162.
Mehra, R. and E.Prescott (1988), ‘The equity risk premium: a solution?’ Journal of Monetary
Economics, 22, 133–136.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
354
S144
CALIBRATION
F.CANOVA
McGratten, E., R.Rogerson and R.Wright (1991), ‘Estimating the stochastic growth model
with household production’, Federal Reserve Bank of Minneapolis, manuscript.
Niederreiter, H. (1988), ‘Quasi Monte Carlo methods for multidimensional numerical
integration’, International Series of Numerical Mathematics, 85, 157–171.
Novales, A. (1990), ‘Solving nonlinear rational expectations models: a stochastic equilibrium
model of interest rates’, Econometrica, 58, 93–111.
Pagan, A. and A.Ullah (1991), ‘Nonparametric estimation’, University of Rochester,
manuscript.
Phillips, P.C.B. (1991), ‘To criticize the critics: an objective Bayesian analysis of stochastic
trends’, Journal of Applied Econometrics, 6(4), 333–354.
Prescott, E. (1991), ‘Real business cycle theories: what have we learned?’ Federal Research
of Minneapolis, Working Paper 486.
Reitz, T. (1988), ‘The equity risk premium: a solution’, Journal of Monetary Economics, 22,
117–132.
Shoven, J. and J.Whalley (1984), ‘Applied general equilibrium models of taxation and
international trade: an introduction and survey’, Journal of Economic Literature, 22, 1007–
1051.
Sims, C. (1984), ‘Solving nonlinear stochastic equilibrium models backward’, University of
Minnesota, manuscript.
Smith, T. (1993), ‘Estimating nonlinear time series models using simulated VAR’, Journal of
Applied Econometrics, 8, s63–s84.
Tanner, M. and W.Wong (1987), ‘The calculation of posterior distributions by data
augmentation’, Journal of the American Statistical Association, 87, 528–550.
Tauchen, G. and R.Hussey (1991), ‘Quadrature based methods for obtaining approximate
solutions to integral equations of nonlinear asset pricing models’, Econometrica, 59, 371–
397.
Watson, M. (1993), ‘Measures of fit for calibrated models’, Journal of Political Economy, 101,
1011–1041.
Weil, P. (1990), ‘The equity premium puzzle and the risk free puzzle’, Journal of Monetary
Economics, 24, 401–421.
Wolf, F. (1986), Meta-Analysis: Quantitative Methods for Research Synthesis, Beverly Hills, CA:
Sage.
© 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin
D.Salyer; individual essays © their authors
CHAPTER 19
355
INTERNATIONAL ECONOMIC REVIEW
Vol. 36, No. 2, May 1995
SENSITIVITY ANALYSIS AND MODEL EVALUATION IN
SIMULATED DYNAMIC GENERAL EQUILIBRIUM
ECONOMIES*
BY FABIO CANOVA1
This paper describes a Monte Carlo procedure to evaluate dynamic nonlinear general
equilibrium macro models. The procedure makes the choice of parameters and the
evaluation of the model less subjective than standard calibration techniques, it
provides more general restrictions than estimation by simulation approaches and
provides a way to conduct global sensitivity analysis for reasonable perturbations of
the parameters. As an illustration the technique is applied to three examples involving
different models and statistics.
1. INTRODUCTION
A growing body of research in the applied macroeconomic literature uses simulation
techniques to derive the time series properties of nonlinear stochastic general equilibrium models, to compare them to real world data and to evaluate policy options (see
e.g. King, Plosser, and Rebelo 1988, or Cooley and Hansen 1990). In implementing
numerical analyses of general equilibrium models, one has to overcome four hurdles.
First, an economy must be specified and functional forms for its primitives selected.
Second, a decision rule for the endogenous variables in terms of the exogenous (and
predetermined) variables and of the parameters must be computed. Third, given the
probability structure of the economy, values for the parameters must be chosen.
Fourth, the closeness of functions of simulated and the actual data must be assessed
in a metric which is relevant to the problem and policy conclusions, if any, should be
drawn.
While models are often specified with an eye to analytical tractability and there has
been progress in developing techniques to numerically approximate unknown
decision rules for the endogenous variables (see e.g. Sims 1984, Coleman 1989,
Novales 1990, Baxter 1991, Tauchen and Hussey 1991, Judd 1992, Marcet 1992 and
the January 1